325 102 3MB
English Pages 360 Year 1999
Adaptive Control Systems
Adaptive Control Systems GANG FENG and ROGELIO LOZANO
OXFORD AUCKLAND BOSTON JOHANNESBURG MELBOURNE NEW DELHI
Newnes An imprint of Butterworth-Heinemann Linacre House, Jordan Hill, Oxford OX2 8DP 225 Wildwood Avenue, Woburn, MA 01801±2041 A division of Reed Educational and Professional Publishing Ltd A member of the Reed Elsevier plc group First published 1999 # Reed Educational and Professional Publishing Ltd 1999 All rights reserved. No part of this publication may be reproduced in any material form (including photocopying or storing in any medium by electronic means and whether or not transiently or incidentally to some other use of this publication) without the written permission of the copyright holder except in accordance with the provisions of the Copyright, Designs and Patents Act 1988 or under the terms of a licence issued by the Copyright Licensing Agency Ltd, 90 Tottenham Court Rd, London, England W1P 9HE. Applications for the copyright holder's written permission to reproduce any part of this publication should be addressed to the publishers. British Library Cataloguing in Publication Data A catalogue record for this book is available from the British Library. ISBN 07506 3996 2 Library of Congress Cataloguing in Publication Data A catalogue record for this book is available from the Library of Congress.
Typeset by David Gregson Associates, Beccles, Suolk Printed in Great Britain by Biddles Ltd, Guildford, Surrey
Contents
List of contributors
ix
Preface
xv
1 Adaptive internal model control 1.1 Introduction 1.2 Internal model control (IMC) schemes: known parameters 1.3 Adaptive internal model control schemes 1.4 Stability and robustness analysis 1.5 Simulation examples 1.6 Concluding remarks References
1 1 2 7 12 18 19 21
2 An algorithm for robust adaptive control with less prior knowledge 2.1 Introduction 2.2 Problem formulation 2.3 Ordinary direct adaptive control with dead zone 2.4 New robust direct adaptive control 2.5 Robust adaptive control with least prior knowledge 2.6 Simulation example 2.7 Conclusions References
23 23 25 27 28 32 36 38 38
3 Adaptive variable structure control 3.1 Introduction 3.2 Problem formulation 3.3 The case of relative degree one 3.4 The case of arbitrary relative degree
41 41 43 46 49
vi
Contents
3.5 Computer simulations 3.6 Conclusion Appendix References 4 Indirect adaptive periodic control 4.1 Introduction 4.2 Problem formulation 4.3 Adaptive control scheme 4.4 Adaptive control law 4.5 Simulations 4.6 Conclusions References
56 59 60 61 63 63 65 68 69 74 78 78
5 Adaptive stabilization of uncertain discrete-time systems via switching control: the method of localization 5.1 Introduction 5.2 Problem statement 5.3 Direct localization principle 5.4 Indirect localization principle 5.5 Simulation examples 5.6 Conclusions Appendix A Appendix B References
80 80 86 89 101 108 112 112 115 116
6 Adaptive nonlinear control: passivation and small gain techniques 6.1 Introduction 6.2 Mathematical preliminaries 6.3 Adaptive passivation 6.4 Small gain-based adaptive control 6.5 Conclusions References
119 119 121 127 138 155 156
7 Active identi®cation for control of discrete-time uncertain nonlinear systems 7.1 Introduction 7.2 Problem formulation 7.3 Active identi®cation 7.4 Finite duration 7.5 Concluding remarks Appendix References
159 160 161 166 178 178 181 182
Contents
vii
8 Optimal adaptive tracking for nonlinear systems 8.1 Introduction 8.2 Problem statement: adaptive tracking 8.3 Adaptive tracking and atclf 's 8.4 Adaptive backstepping 8.5 Inverse optimal adaptive tracking 8.6 Inverse optimality via backstepping 8.7 Design for strict-feedback systems 8.8 Transient performance 8.9 Conclusions Appendix A technical lemma References
184 184 186 188 193 197 202 205 209 210 210 213
9 Stable adaptive systems in the presence of nonlinear parametrization 9.1 Introduction 9.2 Statement of the problem 9.3 Preliminaries 9.4 Stable adaptive NP-systems 9.5 Applications 9.6 Conclusions Appendix Proofs of lemmas References
215 215 218 220 228 236 245 246 258
10 Adaptive inverse for actuator compensation 10.1 Introduction 10.2 Plants with actuator nonlinearities 10.3 Parametrized inverses 10.4 State feedback designs 10.5 Output feedback inverse control 10.6 Output feedback designs 10.7 Designs for multivariable systems 10.8 Designs for nonlinear dynamics 10.9 Concluding remarks References
260 260 262 263 265 269 271 275 281 284 285
11 Stable multi-input multi-output adaptive fuzzy/neural control 11.1 Introduction 11.2 Direct adaptive control 11.3 Indirect adaptive control 11.4 Applications 11.5 Conclusions References
287 287 289 296 299 306 306
viii
Contents
12 Adaptive robust control scheme with an application to PM synchronous motors 12.1 Introduction 12.2 Problem formulation 12.3 Adaptive robust control with -modi®cation 12.4 Application to PM synchronous motors 12.5 Conclusion Appendix A The derivative of 1 Appendix B The derivative of 2 Appendix C The derivative of 3 Appendix D Appendix E De®nitions of 2 , 3 , 2 , and 3
308 309 310 312 317 320 321 322 324 325 326
Index
000
List of contributors
Amit Ailon Department of Electrical and Computer Engineering Ben Gurion University of the Negev Israel
Anuradha M. Annaswamy Adaptive Control Laboratory Department of Mechanical Engineering Massachusetts Institute of Technology Cambridge MA02139 USA
Chiang-Ju Chien Department of Electronic Engineering Hua Fan University Taipei Taiwan ROC
Aniruddha Datta Department of Electrical Engineering Texas A & M University College Station TX 77843±3128, USA
x
List of contributors
Dimitrios Dimogianopoulos Universite de Technologie de Compiegne HEUDIASYC UMR 6599 CNRS-UTC BP 20.529 60200 Compiegne France Gang Feng School of Electical Engineering University of New South Wales Sydney, NSW 2052 Australia Li-Chen Fu Department of Computer Science & Information Engineering National Taiwan University Taipei Taiwan ROC Minyue Fu Department of Electrical and Computer Engineering The University of Newcastle NSW 2308 Australia David J. Hill School of Electrical and Information Engineering Bldg J13 Sydney University New South Wales 2006, Australia Qing-Wei Jia Department of Electrical Engineerng National University of Singapore 10 Kent Ridge Crescent Singapore 119260 Y. A. Jiang School of Electrical Engineering University of New South Wales Sydney NSW 2052 USA
List of contributors
Zhong-Ping Jiang Department of Electrical Engneering University of California Riverside, CA 92521 USA Ioannis Kanellakopoulos UCLA Electrical Engineering Box 951594 Los Angeles, CA 90095±1594, USA Miroslav KrsÏ tic Department of Mechanical Engineering University of Maryland College Park MD 20742, USA T. H. Lee Department of Electrical Engineering National University of Singapore 10 Kent Ridge Crescent Singapore 119260 Zhong-Hua Li Department of Mechanical Engineering University of Maryland College Park MD 20742 USA Ai-Poh Loh Department of Electrical Engineering National University of Singapore Singapore 119260 Rogelio Lozano Universite de Technologie de Compiegne HEUDIASYC UMR 6599 CNRS-UTC BP 20.529 60200 Compiegne France
xi
xii
List of contributors
Richard H. Middleton Department of Electrical and Computer Engineering The University of Newcastle New South Wales 2308, Australia
Raul Ordonez Department of Electrical Engineering The Ohio State University 2015 Neil Avenue Columbus Ohio 43210 USA
Kevin M. Passino Department of Electrical Engineering The Ohio State University 2015 Neil Avenue Columbus OH 43210±1272, USA
Gang Tao Department of Electrical Engineering University of Virginia Charlottesville VA 22903, USA
Lei Xing Department of Electrical Engineeing Texas A&M University College Station TX 77843-3128 USA
J. X. Xu Department of Electrical Engineering National University of Singapore 10 Kent Ridge Crescent Singapore 119260
List of contributors
Jiaxiang Zhao UCLA Electrical Engineering Box 951594 Los Angeles CA 90095-1594 USA Peter V. Zhivoglyadov Department of Electrical and Computer Engineering The University of Newcastle NSW 2308 Australia R. Zmood Department of Communication and Electrical Engineering Royal Melbourne Institute of Technology Melbourne Victoria 3001 Australia
xiii
Preface
Adaptive control has been extensively investigated and developed in both theory and application during the past few decades, and it is still a very active research ®eld. In the earlier stage, most studies in adaptive control concentrated on linear systems. A remarkable development of the adaptive control theory is the resolution of the so-called ideal problem, that is, the proof that several adaptive control systems are globally stable under certain ideal conditions. Then the robustness issues of adaptive control with respect to nonideal conditions such as external disturbances and unmodelled dynamics were addressed which resulted in many dierent robust adaptive control algorithms. These robust algorithms include dead zone, normalization, -modi®cation, 1 modi®cation among many others. At the same time, extensive study has been carried out for reducing a priori knowledge of the systems and improving the transient performance of adaptive control systems. Most recently, adaptive control of nonlinear systems has received great attention and a number of signi®cant results have been obtained. In this book, we have compiled some of the most recent developments of adaptive control for both linear and nonlinear systems from leading world researchers in the ®eld. These include various robust techniques, performance enhancement techniques, techniques with less a priori knowledge, adaptive switching techniques, nonlinear adaptive control techniques and intelligent adaptive control techniques. Each technique described has been developed to provide a practical solution to a real-life problem. This volume will therefore not only advance the ®eld of adaptive control as an area of study, but will also show how the potential of this technology can be realized and oer signi®cant bene®ts. The ®rst contribution in this book is `Adaptive internal model control' by A. Datta and L. Xing. It develops a systematic theory for the design and analysis of adaptive internal model control schemes. The ubiquitous certainty equiva-
xvi
Preface
lence principle of adaptive control is used to combine a robust adaptive law with robust internal model controllers to obtain adaptive internal model control schemes which can be proven to be robustly stable. Speci®c controller structures considered include those of the model reference, partial pole placement, and 2 and 1 optimal control types. The results here not only provide a theoretical basis for analytically justifying some of the reported industrial successes of existing adaptive internal model control schemes but also open up the possibility of synthesizing new ones by simply combining a robust adaptive law with a robust internal model controller structure. The next contribution is `An algorithm for robust direct adaptive control with less prior knowledge' by G. Feng, Y. A. Jiang and R. Zmood. It discusses several approaches to minimizing a priori knowledge required on the unknown plants for robust adaptive control. It takes a discrete time robust direct adaptive control algorithm with a dead zone as an example. It shows that for a class of unmodelled dynamics and bounded disturbances, no knowledge of the parameters of the upper bounding function on the unmodelled dynamics and disturbances is required a priori. Furthermore it shows that a correction procedure can be employed in the least squares estimation algorithm so that no knowledge of the lower bound on the leading coecient of the plant numerator polynomial is required to achieve the singularity free adaptive control law. The global stability and convergence results of the algorithm are established. The next contribution is `Adaptive variable structure control' by C. J. Chiang and Lichen Fu. A uni®ed algorithm is presented to develop the variable structure MRAC for an SISO system with unmodelled dynamics and output measurement noises. The proposed algorithm solves the robustness and performance problem of the traditional MRAC with arbitrary relative degree. It is shown that without any persistent excitation the output tracking error can be driven to zero for relative degree-one plants and driven to a small residual set asymptotically for plants with any higher relative degree. Furthermore, under suitable choice of initial conditions on control parameters, the tracking performance can be improved, which is hardly achievable by the traditional MRAC schemes, especially for plants with uncertainties. The next contribution is `Indirect adaptive periodic control' by D. Dimogianopoulos, R. Lozano and A. Ailon. This new, indirect adaptive control method is based on a lifted representation of the plant which can be stabilized using a simple performant periodic control scheme. The controller parameters computation involves the inverse of the controllability/observability matrix. Potential singularities of this matrix are avoided by means of an appropriate estimates modi®cation. This estimates transformation is linked to the covariance matrix properties and hence it preserves the convergence properties of the original estimates. This modi®cation involves the singular value decomposition of the controllability/observability matrix's estimate. As compared to previous studies in the subject the controller proposed here does
Preface
xvii
not require the frequent introduction of periodic -length sequences of zero inputs. Therefore the new controller is such that the system is almost always operating in closed loop which should lead to better performance characteristics. The next contribution is `Adaptive stabilization of uncertain discrete-time systems via switching control: the method of localization' by P. V. Zhivoglyadov, R. Middleton and M. Fu. It presents a new systematic switching control approach to adaptive stabilization of uncertain discrete-time systems. The approach is based on a method of localization which is conceptually dierent from supervisory adaptive control schemes and other existing switching control schemes. The proposed approach allows for slow parameter drifting, infrequent large parameter jumps and unknown bound on exogenous disturbances. The unique feature of the localization-based switching adaptive control proposed here is its rapid model falsi®cation capability. In the LTI case this is manifested in the ability of the switching controller to quickly converge to a suitable stabilizing controller. It is believed that the approach is applicable to a wide class of linear time invariant and time-varying systems with good transient performance. The next contribution is `Adaptive nonlinear control: passivation and small gain techniques' by Z. P. Jiang and D. Hill. It proposes methods to systematically design stabilizing adaptive controllers for new classes of nonlinear systems by using passivation and small gain techniques. It is shown that for a class of linearly parametrized nonlinear systems with only unknown parameters, the concept of adaptive passivation can be used to unify and extend most of the known adaptive nonlinear control algorithms based on Lyapunov methods. A novel recursive robust adaptive control method by means of backstepping and small gain techniques is also developed to generate a new class of adaptive nonlinear controllers with robustness to nonlinear unmodelled dynamics. The next contribution is `Active identi®cation for control of discrete-time uncertain nonlinear systems' by J. Zhao and I. Kanellakopoulos. A novel approach is proposed to remove the restrictive growth conditions of the nonlinearities and to yield global stability and tracking for systems that can be transformed into an output-feedback canonical form. The main novelties of the design are (i) the temporal and algorithmic separation of the parameter estimation task from the control task and (ii) the development of an active identi®cation procedure, which uses the control input to actively drive the system state to points in the state space that allow the orthogonalized projection estimator to acquire all the necessary information about the unknown parameters. It is proved that the proposed algorithm guarantees complete identi®cation in a ®nite time interval and global stability and tracking.
xviii
Preface
The next contribution is `Optimal adaptive tracking for nonlinear systems' by M. KrsÏ tic and Z. H. Li. In this chapter an `inverse optimal' adaptive tracking problem for nonlinear systems with unknown parameters is de®ned and solved. The basis of the proposed method is an adaptive tracking control Lyapunov function (atclf) whose existence guarantees the solvability of the inverse optimal problem. The controllers designed are not of certainty equivalence type. Even in the linear case they would not be a result of solving a Riccati equation for a given value of the parameter estimate. Inverse optimality is combined with backstepping to design a new class of adaptive controllers for strict-feedback systems. These controllers solve a problem left open in the previous adaptive backstepping designs ± getting transient performance bounds that include an estimate of control eort. The next contribution is `Stable adaptive systems in the presence of nonlinear parameterization' by A. M. Annaswamy and A. P. Loh. This chapter addresses the problem of adaptive control when the unknown parameters occur nonlinearly in a dynamic system. The traditional approach used in linearly parameterized systems employs a gradient-search principle in estimating the unknown parameters. Such an approach is not sucient for nonlinearly parametrized systems. Instead, a new algorithm based on a min±max optimization scheme is developed to address nonlinearly parametrized adaptive systems. It is shown that this algorithm results in globally stable closed loop systems when the states of the plant are accessible for measurement. The next contribution is `Adaptive inverse for actuator compensation' by G. Tao. A general adaptive inverse approach is developed for control of plants with actuator imperfections caused by nonsmooth nonlinearities such as deadzone, backlash, hysteresis and other piecewise-linear characteristics. An adaptive inverse is employed for cancelling the eect of an unknown actuator nonlinearity, and a linear feedback control law is used for controlling the dynamics of a known linear or smooth nonlinear part following the actuator nonlinearity. State feedback and output feedback control designs are presented which all lead to linearly parametrized error models suitable for the development of adaptive laws to update the inverse parameters. This approach suggests that control systems with commonly used linear or nonlinear feedback controllers such as those with an LQ, model reference, PID, pole placement or other dynamic compensation design can be combined with an adaptive inverse for improving system tracking performance despite the presence of actuator imperfections. The next contribution is `Stable multi-input multi-output adaptive fuzzy/ neural control' by R. OrdoÂnÄez and K. Passino. In this chapter, stable direct and indirect adaptive controllers are presented which use Takagi±Sugeno fuzzy systems, conventional fuzzy systems, or a class of neural networks to provide asymptotic tracking of a reference signal vector for a class of continuous time multi-input multi-output (MIMO) square nonlinear plants with poorly under-
Preface
xix
stood dynamics. The direct adaptive scheme allows for the inclusion of a priori knowledge about the control input in terms of exact mathematical equations or linguistics, while the indirect adaptive controller permits the explicit use of equations to represent portions of the plant dynamics. It is shown that with or without such knowledge the adaptive schemes can `learn' how to control the plant, provide for bounded internal signals, and achieve asymptotically stable tracking of the reference inputs. No initialization condition needs to be imposed on the controllers, and convergence of the tracking error to zero is guaranteed. The ®nal contribution is `Adaptive robust control scheme with an application to PM synchronous motors' by J. X. Xu, Q. W. Jia and T. H. Lee. A new, adaptive, robust control scheme for a class of nonlinear uncertain dynamical systems is presented. To reduce the robust control gain and widen the application scope of adaptive techniques, the system uncertainties are classi®ed into two dierent categories: the structured and nonstructured uncertainties with partially known bounding functions. The structured uncertainty is estimated with adaptation and compensated. Meanwhile, the adaptive robust method is applied to deal with the non-structured uncertainty by estimating unknown parameters in the upper bounding function. It is shown that the new control scheme guarantees the uniform boundedness of the system and assures the tracking error entering an arbitrarily designated zone in a ®nite time. The eectiveness of the proposed method is demonstrated by the application to PM synchronous motors.
1
Adaptive internal model control A. Datta and L. Xing
Abstract This chapter develops a systematic theory for the design and analysis of adaptive internal model control schemes. The principal motivation stems from the fact that despite the reported industrial successes of adaptive internal model control schemes, there currently does not exist a design methodology capable of providing theoretical guarantess of stability and robustness. The ubiquitous certainty equivalence principle of adaptive control is used to combine a robust adaptive law with robust internal model controllers to obtain adaptive internal model control schemes which can be proven to be robustly stable. Speci®c controller structures considered include those of the model reference, `partial' pole placement, and 2 and 1 optimal control types. The results here not only provide a theoretical basis for analytically justifying some of the reported industrial successes of existing adaptive internal model control schemes but also open up the possibility of synthesizing new ones by simply combining a robust adaptive law with a robust internal model controller structure.
1.1
Introduction
Internal model control (IMC) schemes, where the controller implementation includes an model of the plant, continue to enjoy widespread popularity in industrial process control applications [1]. Such schemes can guarantee internal stability for only plants; since most plants encountered in process control are anyway open loop stable, this really does not impose any signi®cant restriction.
2
Adaptive internal model control
As already mentioned, the main feature of IMC is that its implementation requires an explicit model of the plant to be used as part of the controller. When the plant itself happens to be unknown, or the plant parameters vary slowly with time due to ageing, no such model is directly available a priori and one has to resort to identi®cation techniques to come up with an appropriate plant model on-line. Several empirical studies, e.g. [2], [3] have demonstrated the feasibility of such an approach. However, what is, by and large, lacking in the process control literature is the availability of results with solid theoretical guarantees of stability and performance. Motivated by this fact, in [4], [5], we presented designs of adaptive IMC schemes with provable guarantees of stability and robustness. The scheme in [4] involved on-line adaptation of only the internal model while in [5], in addition to adapting the internal model on-line, the IMC parameter was chosen in a certainty equivalence fashion to pointwise optimize an 2 performance index. In this chapter, it is shown that the approach of [5] can be adapted to design and analyse a class of adaptive 1 optimal control schemes that are likely to arise in process control applications. This class speci®cally consists of those 1 norm minimization problems that involve only . Additionally, we reinterpret the scheme of [4] as an adaptive `partial' poleplacement control scheme and consider the design and analysis of a model reference adaptive control scheme based on the IMC structure. In other words, this chapter considers the design and analysis of popular adaptive control schemes from the literature within the context of the IMC con®guration. A single, uni®ed, analysis procedure, applicable to each of the schemes considered, is also presented. The chapter is organized as follows. In Section 1.2, we present several nonadaptive control schemes utilizing the IMC con®guration. Their adaptive certainty equivalence versions are presented in Section 1.3. A uni®ed stability and robustness analysis encompassing all of the schemes of Section 1.3 is presented in Section 1.4. In Section 1.5, we present simulation examples to demonstrate the ecacy of our adaptive IMC designs. Section 1.6 concludes the chapter by summarizing the main results and outlining their expected signi®cance.
1.2
Internal model control (IMC) schemes: known parameters
In this section, we present several nonadaptive control schemes utilizing the IMC structure. To this end, we consider the IMC con®guration for a stable plant
as shown in Figure 1.1. The IMC controller consists of a stable `IMC parameter'
and a model of the plant which is usually referred to as the `internal model'. It can be shown [1, 4] that if the plant
is stable and
Adaptive Control Systems
3
Internal model controller r
Internal model
Figure 1.1 The IMC con®guration
the internal model is an exact replica of the plant, then the stability of the IMC parameter is equivalent to the internal stability of the con®guration in Figure 1.1. Indeed, the IMC parameter is really the Youla parameter [6] that appears in a special case of the YJBK parametrization of all stabilizing controllers [4]. Because of this, internal stability is assured as long as
is chosen to be any stable rational transfer function. We now show that dierent choices of stable
lead to some familiar control schemes. 1.2.1 Partial pole placement control From Figure 1.1, it is clear that if the internal model is an exact replica of the plant, then there is no feedback signal in the loop. Consequently the poles of the closed loop system are made up of the open loop poles of the plant and the poles of the IMC parameter
. Thus, in this case, a `complete' pole placement as in traditional pole placement control schemes is not possible. Instead, one can only choose the poles of the IMC parameter
to be in some desired locations in the left half plane while leaving the remaining poles at the plant open loop pole locations. Such a control scheme, where
is chosen to inject an additional set of poles at some desired locations in the complex plane, is referred to as `partial' pole placement. 1.2.2 Model reference control The objective in model reference control is to design a dierentiator-free controller so that the output of the controlled plant
asymptotically tracks the output of a stable reference model
for piecewise continuous reference input signals
. In order to meet the control objective, we make the following assumptions which are by now standard in the model reference control literature: (M1) The plant
is minimum phase; and (M2) The relative degree of the reference model transfer function
is greater than or equal to that of the plant transfer function
.
Adaptive internal model control
4
Assumption (M1) above is necessary for ensuring internal stability since satisfaction of the model reference control objective requires cancellation of the plant zeros. Assumption (M2), on the other hand, permits the design of a dierentiator-free controller to meet the control objective. If assumptions (M1) and (M2) are satis®ed, it is easy to verify from Figure 1.1 that the choice
1
1:1
for the IMC parameter guarantees the satisfaction of the model reference control objective in the ideal case, i.e. in the absence of plant modelling errors. 1.2.3 H2 optimal control In 2 optimal control, one chooses
to minimize the 2 norm of the tracking error provided 2 2 . From Figure 1.1, we obtain
) )
1
1
2
k1
0
k2 2 (using Parseval's Theorem)
where
is the Laplace transform of
and k
k2 denotes the standard 2 norm. Thus the mathematical problem of interest here is to choose
to minimize k1
k2 . The following theorem gives the analytical expression for the minimizing
. The detailed derivation can be found in [1]. Theorem 2.1 Let
be the stable plant to be controlled and let
be the Laplace Transform of the external input signal
1. Suppose that
has no poles in the open right half plane2 and that there exists at least one choice, say 0
, of the stable IMC parameter
such that 1
0
is stable3. Let 1 ; 2 ; ; be the open right half plane zeros of
and de®ne the Blaschke product4
1
2
1
2
so that
can be rewritten as
1
For the sake of simplicity, both
and
are assumed to be rational transfer functions. The theorem statement can be appropriately modi®ed for the case where
and/or
contain all-pass time delay factors [1] 2 This assumption is reasonable since otherwise the external input would be unbounded. 3 The ®nal construction of the 2 optimal controller serves as proof for the existence of a 0
with such properties. 4 Here
denotes complex conjugation.
Adaptive Control Systems
5
where
is minimum phase. Similarly, let 1 ; 2 ; ; be the open right half plane zeros of
and de®ne the Blashcke product
1
2
1
2
so that
can be rewritten as
where
is minimum phase. Then the
which minimizes k1
k2 is given by
1
1
1
1:2
where denotes that after a partial fraction expansion, the terms corresponding to the poles of 1
are removed.
Remark 2.1 The optimal
de®ned in (1.2) is usually improper. So it is customary to make
proper by introducing sucient high frequency attenuation via what is called the `IMC Filter'
[1]. Instead of the optimal
in (1.2), the
to be implemented is given by
1
1
1
1:3
where
is the stable IMC ®lter. The design of the IMC ®lter for 2 optimal control depends on the choice of the input
. Although this design is carried out in a somewhat ad hoc fashion, care is taken to ensure that the original asymptotic tracking properties of the controller are preserved. This is because otherwise 1
may no longer be a function in 2 . As a speci®c example, suppose that the system is of Type 1.5 Then, a possible choice for the IMC ®lter to ensure retention of asymptotic tracking properties is 1
is chosen to be a large enough positive , > 0 where
1
integer to make
proper. As shown in [1], the parameter represents a trade-o between tracking performance and robustness to modelling errors. 1.2.4 H1 optimal control The sensitivity function
and the complementary sensitivity function
for the IMC con®guration in Figure 1.1 are given by
1
and
respectively [1]. Since the plant
is open loop stable, it follows that the 1 norm of the complementary sensitivity function
can 1 be made by simply choosing
and letting tend to 5
Other system types can also be handled as in [1].
6
Adaptive internal model control
in®nity. Thus minimizing the 1 norm of
does not make much sense since the in®mum value of zero is unattainable. On the other hand, if we consider the weighted sensitivity minimization problem where we seek to minimize k
k1 for some stable, minimum phase, rational weighting transfer function
, then we have an interesting 1 minimization problem, i.e. choose a stable
to minimize k
1
k1 .The solution to this problem depends on the number of open right half plane zeros of the plant
and involves the use of Nevanlinna±Pick interpolation when the plant
has more than one right half plane zero [7]. However, when the plant has only one right half plane zero
1 and none on the imaginary axis, there is only one interpolation constraint and the closed form solution is given by [7]
1
1 1
1:4
Fortunately, this case covers a large number of process control applications where plants are typically modelled as minimum phase ®rst or second order transfer functions with time delays. Since approximating a delay using a ®rst order Pade approximation introduces one right half plane zero, the resulting rational approximation will satisfy the one right half plane zero assumption. Remark 2.2 As in the case of 2 optimal control, the optimal
de®ned by (1.4) is usually improper. This situation can be handled as in Remark 2.1 so that the
to be implemented becomes
1
1:5
1 1
where
is a stable IMC ®lter. In this case, however, there is more freedom in the choice of
since the 1 optimal controller (1.4) does not necessarily guarantee any asymptotic tracking properties to start with. 1.2.5 Robustness to uncertainties (small gain theorem) In the next section, we will be combining the above schemes with a robust adaptive law to obtain adaptive IMC schemes. If the above IMC schemes are unable to tolerate uncertainty in the case where all the plant parameters are known, then there is little or no hope that certainty equivalence designs based on them will do any better when additionally the plant parameters are unknown and have to be estimated using an adaptive law. Accordingly, we now establish the robustness of the nonadaptive IMC schemes to the presence of plant modelling errors. Without any loss of generality let us suppose that the uncertainty is of the multiplicative type, i.e.
0
1
1:6
Adaptive Control Systems
7
where 0
is the modelled part of the plant and
is a stable multiplicative uncertainty such that 0
is strictly proper. Then we can state the following robustness result which follows immediately from the small gain theorem [8]. A detailed proof can also be found in [1]. Theorem 2.2 Suppose 0
and
are stable transfer functions so that the IMC con®guration in Figure 1.1 is stable for
0
. Then the IMC con®guration with the actual plant given by (1.6) is still stable provided 1 . 2 0; where k0
k1
1.3
Adaptive internal model control schemes
In order to implement the IMC-based controllers of the last section, the plant must be known a priori so that the `internal model' can be designed and the IMC parameter
calculated. When the plant itself is unknown, the IMCbased controllers cannot be implemented. In this case, the natural approach to follow is to retain the same controller structure as in Figure 1.1, with the internal model being adapted on-line based on some kind of parameter estimation mechanism, and the IMC parameter
being updated pointwise using one of the above control laws. This is the standard certainty equivalence approach of adaptive control and results in what are called adaptive internal model control schemes. Although such adaptive IMC schemes have been empirically studied inthe literature, e.g. [2, 3], our objective here is to develop adaptive IMC schemes with provable guarantees of stability and robustness. To this end, we assume that the stable plant to be controlled is described by
0
1
; 0
>0
1:7
where 0
is a monic Hurwitz polynomial of degree ; 0
is a polynomial 0
represents the modelled part of the plant; and of degree with < ; 0
0
is a stable multiplicative uncertainty such that
is strictly 0
proper. We next present the design of the robust adaptive law which is carried out using a standard approach from the robust adaptive control literature [9]. 1.3.1 Design of the robust adaptive law We start with the plant equation
0
1
; 0
> 0
1:8
8
Adaptive internal model control
where , are the plant input and output signals. This equation can be rewritten as 0
0
0
1 , where
is an arbitrary, monic, Hurwitz
polynomial of degree , we obtain
Filtering both sides by
0
0
0
1:9
The above equation can be rewritten as
where
1 ;
0
; 2
and
2 ;
and
1 ,
0
1:10
2
are vectors containing the coecients of 1
, respectively; 1 ; 2 ; 1
1
1 ; 2 ; ; 1
; 1 ; ; 1 D
0
1:11
Equation (1.10) is exactly in the form of the linear parametric model with modelling error for which a large class of robust adaptive laws can be developed. In particular, using the gradient method with normalization and parameter projection, we obtain the following robust adaptive law [9] ";
0 2 C "
1:12
1:13
2
2 1 2 ;
1:14
2
0 2 2 ;
1:15
0 0
1:16
where > 0 is an adaptive gain; C is a known compact convex set containing ; is the standard projection operator which guarantees that the parameter estimate
does not exit the set C and 0 > 0 is a constant chosen so that 0 . This choice of , of course,
, 1 are analytic in R
2 necessitates some a priori knowledge about the stability margin of the
Adaptive Control Systems
9
unmodelled dynamics, an assumption which has by now become fairly standard in the robust adaptive control literature [9]. The robust adaptive IMC schemes are obtained by replacing the internal model in Figure 1.1 by that obtained from equation (1.14), and the IMC parameters
by timevarying operators which implement the certainty equivalence versions of the controller structures considered in the last section. The design of these certainty equivalence controllers is discussed next. 1.3.2 Certainty equivalence control laws We ®rst outline the steps involved in designing a general certainty equivalence adaptive IMC scheme. Thereafter, additional simpli®cations or complexities that result from the use of a particular control law will be discussed. . ! First use the parameter estimate
obtained from the robust adaptive law (1.12)±(1.16) to generate estimates of the numerator and denominator polynomials for the modelled part of the plant6
0
; 2
0
;
1
1
0
; , calculate the appro. "! Using the frozen time plant
; 0
; priate
; using the results developed in Section 1.2. as
;
; where
; and
; are . #! Express
;
; time-varying polynomials with
; $ . . %! Choose 1
to be an arbitrary monic Hurwitz polynomial of degree equal to that of
; , and let this degree be denoted by . . &! The certainty equivalence control law is given by '
1
'
1
1
"2
1:17
where '
is the vector of coecients of 1
; ; '
is the vector of coecients of
; ;
; 1 ; ; 1 and 1
1 ; 2 ; ; 1 . The robust adaptive IMC scheme resulting from combining the control law (1.17) with the robust adaptive law (1.12)±(1.16) is schematically depicted in 6
In the rest of this chapter, the `hats' denote the time varying polynomials/frozen time `transfer functions' that result from replacing the time-invariant coecients of a `hat-free' polynomial/transfer function by their corresponding time-varying values obtained from adaptation and/or certainty equivalence control.
10
Adaptive internal model control Regressor generating block
L1
'
'
1
L1
" 2
Figure 1.2 Robust adaptive IMC scheme
Figure 1.2. We now proceed to discuss the simpli®cations or additional complexities that result from the use of each of the controller structures presented in Section 1.2. 1.3.2.1
Partial adaptive pole placement
In this case, the design of the IMC parameter does not depend on the estimated plant. Indeed,
is a ®xed stable transfer function and not a time-varying operator so that we essentially recover the scheme presented in [4]. Consequently, this scheme admits a simpler stability analysis as in [4] although the general analysis procedure to be presented in the next section is also applicable. 1.3.2.2
Model reference adaptive control
in Step 2 of the certainty In this case from (1.1), we see that the
; equivalence design becomes
; 1
;
1:18
Our stability analysis to be presented in the next section is based on results in the area of slowly time-varying systems. In order for these results to be be pointwise stable and applicable, it is required that the operator
; also that the degree of
; in Step 3 of the certainty equivalence design not change with time. These two requirements can be satis®ed as follows:
can be guaranteed by ensuring that the . The pointwise stability of
; frozen time estimated plant is minimum phase, i.e. 0
; is Hurwitz stable for every ®xed . To guarantee such a property for 0
; , the projection set C in (1.12)±(1.16) is chosen so that 8 2 C , the corresponding 0
2
is Hurwitz stable. By restricting C to be a subset of a Cartesian product of closed intervals, results from Kharitonov Theory [10] can be used to ensure that C satis®es such a requirement. Also, when the
Adaptive Control Systems
11
projection set C cannot be speci®ed as a single convex set, results from ( )( $ using a ®nite number of convex sets [11] can be used.
; can also be rendered time invariant by ensuring that . The degree of the leading coecient of 0
; is not allowed to pass through zero. This feature can be built into the adaptive law by assuming some knowledge about the sign and a lower bound on the absolute value of the leading coecient of 0
. Projection techniques, appropriately utilizing this knowledge, are by now standard in the adaptive control literature [12]. We will therefore assume that for IMC-based model reference adaptive control, the set C has been suitably chosen to guarantee that the estimate
obtained from (1.12)±(1.16) actually satis®es both of the properties mentioned above. 1.3.2.3 Adaptive H2 optimal control
is obtained by substituting 1
; , 1
; into the rightIn this case,
; and hand side of (1.3) where
; is the minimum phase portion of
;
; is the Blaschke product containing the open right-half plane zeros of is given by 0
; . Thus
; 1
; 1
1
;
;
1:19
where denotes that after a partial fraction expansion, the terms corresponding to the poles of 1
; are removed, and
is an IMC ®lter used to be proper. As will be seen in the next section, speci®cally to force
;
; in Step 3 of the certainty equivalence design Lemma 4.1, the degree of can be kept constant $ $ *
provided the leading coecient of 0
; is not allowed to pass through zero. Additionally 0
; should not have any zeros on the imaginary axis. A parameter projection modi®cation, as in the case of model reference adaptive control, can be incorporated into the adaptive law (1.12)±(1.16) to guarantee both of these properties. 1.3.2.4 Adaptive H1 optimal control
is obtained by substituting
; into the right-hand side of In this case,
; (1.5), i.e.
1 1
;
1:20
; 1
where 1 is the open right half plane zero of 0
; and
is the IMC ®lter. Since (1.20) assumes the presence of only one open right half plane zero, the estimated polynomial 0
; must have only one open right half plane zero and none on the imaginary axis. Additionally the leading coecient of 0
;
; in Step should not be allowed to pass through zero so that the degree of
12
Adaptive internal model control
3 of the certainty equivalence design can be kept ®xed using a single ®xed
. Once again, both of these properties can be guaranteed by the adaptive law by appropriately choosing the set C . Remark 3.1 The actual construction of the sets C for adaptive model reference, adaptive 2 and adaptive 1 optimal control may not be straightforward especially for higher order plants. However, this is a wellknown problem that arises in any certainty equivalence control scheme based on the estimated plant and is really not a drawback associated with the IMC design methodology. Although from time to time a lot of possible solutions to this problem have been proposed in the adaptive literature, it would be fair to say that, by and large, no satisfactory solution is currently available.
1.4
Stability and robustness analysis
Before embarking on the stability and robustness analysis for the adaptive IMC schemes just proposed, we ®rst introduce some de®nitions [9, 4] and state and prove two lemmas which play a pivotal role in the subsequent analysis. De®nition 4.1 For any signal 0; 1 ! , denotes the truncation of to the interval 0; and is de®ned as
if t
1:21 0 otherwise De®nition 4.2 is de®ned as
For any signal 0; 1 ! , and for any 0, 0, k k2 D
k k2
0
12
1:22
The k
k2 represents the exponentially weighted 2 norm of the signal truncated to 0; . When 0 and 1, k
: k2 becomes the usual 2 norm and will be denoted by k:k2 . It can be shown that k:k2 satis®es the usual properties of the vector norm. De®nition 4.3 Consider the signals 0; 1 ! , 0; 1 ! and the set S
0; 1 ! j
for some 0 and 8 ; 0. We say that is -small in the mean if 2
. Lemma 4.1 In each of the adaptive IMC schemes presented in the last section, the degree of
; in Step 3 of the certainty equivalence design can be made
Adaptive Control Systems
13
time invariant. Furthermore, for the adaptive 2 and 1 designs, this can be done using a single ®xed
. + The proof of this lemma is relatively straightforward except in the case of adaptive 2 optimal control. Accordingly, we ®rst discuss the simpler cases before giving a detailed treatment of the more involved one. For adaptive partial pole placement, the time invariance of the degree of
; follows trivially from the fact that the IMC parameter in this case is time invariant. For model reference adaptive control, the fact that the leading coecient of 0
; is not allowed to pass through zero guarantees that the
; is time invariant. Finally, for adaptive 1 optimal control, degree of the result follows from the fact that the leading coecient of 0
; is not allowed to pass through zero. We now present the detailed proof for the case of adaptive 2 optimal control. Let ; be the degrees of the denominator and numerator in (1.19), polynomials respectively of
. Then, in the expression for
;
th order polynomial while 1
it is clear that 1
; th order polynomial
th order polynomial
1th order polynomial . Also 1
;
th order polynomial
th order polynomial where , strict inequality being attained when some of the poles of
coincide with some of the stable zeros of 1
; . Moreover, in any case, the th order denominator polynomial of 1
;
is a factor of given in the th order numerator polynomial of 1
. Thus for the
; (1.19), if we disregard
, then the degree of the numerator polynomial is
1 while that of the denominator polynomial is 4 1. Hence, the degree of
; in Step 3 of the certainty equivalence design can be kept ®xed at
1, and this can be achieved with $ *
of relative degree 1, provided that the leading coecient of 0
; is appropriately constrained. Remark 4.1 Lemma 4.1 tells us that the degree of each of the certainty equivalence controllers presented in the last section can be made time invariant. This is important because, as we will see, it makes it possible to carry out the analysis using standard state-space results on slowly time-varying systems. Lemma 4.2 At any ®xed time , the coecients of
; ,
; , and hence the vectors '
, '
, are continuous functions of the estimate
. + Once again, the proof of this lemma is relatively straightforward except in the case of adaptive 2 optimal control. Accordingly, we ®rst discuss the simpler cases before giving a detailed treatment of the more involved one.
Adaptive internal model control
14
For the case of adaptive partial pole placement control, the continuity follows trivially from the fact that the IMC parameter is independent of
. For model reference adaptive control, the continuity is immediate from (1.18) and the fact that the leading coecient of 0
; is not allowed to pass through zero. Finally for adaptive 1 optimal control, we note that the right half plane zero 1 of 0
; is a continuous function of
. This is a consequence of the fact that the degree of 0
; cannot drop since its leading coecient is not allowed to pass through zero. The desired continuity now follows from (1.20). We now present the detailed proof for the 2 optimal control case. Since the leading coecient of 0
; has been constrained so as not to pass through zero then, for any ®xed , the roots of 0
; are continuous functions of
. Hence, it follows that the coecients of the numerator and denominator 1 are continuous functions of polynomials of
; 1
;
; 1
. Moreover,
;
is the sum of the residues of
; 1
at the poles of
, which clearly depends continuously on
(through the factor
; 1 ). Since
is ®xed and independent of , it follows from (1.19) that the coecients of
; ,
; depend continuously on
. Remark 4.2 Lemma 4.2 is important because it allows one to translate slow variation of the estimated parameter vector
to slow variation of the controller parameters. Since the stability and robustness proofs of most adaptive schemes rely on results from the stability of slowly time-varying systems, establishing continuity of the controller parameters as a function of the estimated plant parameters (which are known to vary slowly) is a crucial ingredient of the analysis. The following theorem describes the stability and robustness properties of the adaptive IMC schemes presented in this chapter. Theorem 4.1 Consider the plant (1.8) subject to the robust adaptive IMC control law (1.12)±(1.16), (1.17), where (1.17) corresponds to any one of the adaptive IMC schemes considered in the last section and
is a bounded external signal. Then, 9 > 0 such that 8 2 0; , all the signals in the 2 2
closed loop system are uniformly bounded and the error 2 2 for 7 some > 0 7
In the rest of this chapter, `c' is the generic symbol for a positive constant. The exact value of such a constant can be determined (for a quantitative robustness result) as in [13, 9]. However, for the qualitative presentation here, the exact values of these constants are not important.
Adaptive Control Systems
15
+ The proof is obtained by combining the properties of the robust adaptive law (1.12)±(1.16) with the properties of the IMC-based controller structure. We ®rst analyse the properties of the adaptive law. From (1.10), (1.13) and (1.14), we obtain
; 2
X
1:23
Consider the positive de®nite function ,
2
Then, along the solution of (1.12), it can be shown that [9] , " " "2
(using (1.23)
1 2 2 1 2 2 " 2 2 2
(completing the squares)
1:24
From (1.11), (1.15), (1.16), using Lemma 2.1 (Equation (7)) in [4], it follows 2 1 . 2 1 . Now, the parameter projection guarantees that ,; ; that Hence integrating both sides of (1.24) from to , we obtain 2 2
" 2 2 Also from (1.12) j"j jj
1:25 jj From the de®nition of , it follows using Lemma 2.1 (equation (7)) in [4] that 2 2
2 1 , which in turn implies that 2 2 . This completes the analysis of the properties of the robust adaptive law. To complete the stability proof, we now turn to the properties of the IMC-based controller structure. The certainty equivalence control law (1.17) can be rewritten as 1 1
1
'
1
1
1
1
"2
where 1
, 2
, ;
are the time-varying coecients of
; . 1 1 ; 2 ; ; ; -X1 ; 2 ; ; , 1 De®ning 1
1
1
16
Adaptive internal model control
the above equation can be rewritten as
where
- .
- '
1
0 1 0
0 0 1 0
D .
1
0
0 D
0 1
2
1:26
0 0
1
Since the time-varying polynomial
; is pointwise Hurwitz, it follows that for any * , the eigenvalues of .
are in the open left half plane. Moreover, since the coecients of
; are continuous functions of
(Lemma 4.2) and
2 C , a compact set, it follows that 9 > 0 such that f
.
g
8 0
and
1; 2; ;
The continuity of the elements of .
with respect to
and the fact that 2 2
2 2
2 S together imply that .
2 S 2 . Hence, using the fact 2 2 1 , it follows from Lemma 3.1 in [9] that 9 1 > 0 such that that 8 2 0; 1 , the equilibrium state 0 of .
is exponentially stable, i.e. there exist 0 ; 0 > 0 such that the state transition matrix
; corresponding to the homogeneous part of (1.26) satis®es k
; k 0
0
8
1:27
1
, it is easy to see that the control input can be From the identity 1
rewritten as
/
- '
"2
1:28 1
where v
; 1 1
; ; 1 1
and
1
1 1
Adaptive Control Systems
Also, using (1.28) in the plant equation (1.8), we obtain 0
2 1
v
- '
0
1
17
1:29
Now let 2
0; min0 ; 0 be chosen such that 0
, 1
are analytic in , and de®ne the ®ctitious normalizing signal +
by R 2
1:30 +
1:0 k k2 k k2 As in [9], we take truncated exponentially weighted norms on both sides of (1.28), (1.29) and make use of Lemma 3.3 in [9] and Lemma 2.1 (equation (6)) in [4], while observing that v
, '
,
2 1 , to obtain k k2 k
"2 k2
1:31
k k2 k
"2 k2
1:32
which together with (1.30) imply that +
k
"2 k2 Now squaring both sides of (1.33) we obtain 2 +
"2 2 2+
1:33
(since
+
0
) 2+
0
"2 2 "
2
2
(using the Bellman-Gronwall lemma [8]) 2 2
Since " 2 and is bounded, it follows using Lemma 2.2 in [4] that 2 9 2
0; 1 such that 8 2 0; , + 2 1 , which in turn implies that , are bounded, it follows that , 2 1 . Thus 2 1 . Since "2 is also bounded so that from (1.26), we obtain - 2 1 . From (1.28), (1.29), we can now conclude that , 2 1 . This establishes the boundedness of all the closed loop signals in the adaptive IMC scheme. Since 2 2
2 2
, it follows that 2 2 as "2 and " 2 , 2 1 2 claimed and, therefore, the proof is complete. Remark 4.3 The robust adaptive IMC schemes of this chapter recover the performance properties of the ideal case if the modelling error disappears, i.e. we can show that if 0 then ! 0 as ! 1. This can be established using standard arguments from the robust adaptive control literature, and is a
18
Adaptive internal model control
consequence of the use of parameter projection as the robustifying modi®cation in the adaptive law [9]. An alternative robustifying modi®cation which can guarantee a similar property is the switching- modi®cation [14].
1.5
Simulation examples
In this section, we present some simulation examples to demonstrate the ecacy of the adaptive IMC schemes proposed. We ®rst consider the plant (1.7) with 0
2, 0
2 1, 1 and 0:01. Choosing 0 0:1, 1,
2 2 2, 3 1 and imC 5:0; 5:0 4:0; 4:0 0:1; 6:0 6:0; 6:0,
4 plementing the adaptive partial pole placement control scheme (1.12)±(1.16), (1.17), with
0 1:0; 2:0; 3:0; 1:0 and all other initial conditions set to zero, we obtained the plots in Figure 1.3 for
1:0 and
sin
0:2. 2 quite well. From these plots, it is clear that
tracks 2
1
4 Let us now consider the design of an adaptive model reference control scheme for the same plant where the reference model is given by 1
2 . The adaptive law (1.12)±(1.16) must now guarantee 2 1 that the estimated plant is pointwise minimum phase, to ensure which, we now choose the set C as C 5:0; 5:0 4:0; 4:0 0:1; 6:0 0:1; 6:0. All the other design parameters are exactly the same as before except that now (1.17) implements the IMC control law (1.18) and 1
3 22 2 2. The resulting plots are shown in Figure 1.4 for
1:0 and
sin
0:2. From these plots, it is clear that the adaptive IMC scheme does achieve model following. The modelled part of the plant we have considered so far is minimum phase which would not lead to an interesting 2 or 1 optimal control problem. Thus, for 2 and 1 optimal control, we consider the plant (1.7) with 1 and 0:01. Choosing 0
1, 0
2 3 2, 3 0 0:1, 1,
2 2 2, 1
2 2 2, C 5:0; 5:0 1 4:0; 4:0 6:0; 0:1 6:0; 6:0,
and implementing the
12 adaptive 2 optimal control scheme (1.12)±(1.16), (1.17), with
0 2:0; 2:0; 2:0; 2:0 and all other initial conditions set to zero, we obtained the plot shown in Figure 1.5. From Figure 1.5, it is clear that
Adaptive Control Systems 1
1
y−system ÐÐ y -system output output ± ± desired ± desiredoutput output r(t)=1 r
t 1
0.8
0.6
0.7
0.4
0.6
0.2
0.5
0
0.4
−0.2
0.3
−0.4
0.2
−0.6
0.1
−0.8
0
50 100 time t (second)
time t (second)
ÐÐ y -system output output y−system ± ± desired ± desiredoutput output r(t)=sin(0.2t) r
t sin
0:2t
0.8
output value
output value
0.9
0
19
150
−1
0
50
100
150
time t (second) time t (second)
Figure 1.3 PPAC IMC simulation
asymptotically tracks
quite well. Note that the projection set C here has been chosen to ensure that the degree of 0
; does not drop. Finally, we simulated an 1 optimal controller for the same plant used for 0:01 and the set the 2 design.The weighting
was chosen as
0:01 C was taken as C 5:0; 5:0 4:0; 4:0 6:0; 0:1 0:1; 6:0. This choice of C ensures that the estimated plant has one and only one right half plane zero. Keeping all the other design parameters the same as in the 2 optimal control case and choosing
1:0 and
0:8 sin
0:2, we obtained the plots shown in Figure 1.6. From these plots, we see that the adaptive 1 -optimal controller does produce reasonably good tracking.
1.6
Concluding remarks
In this chapter, we have presented a general systematic theory for the design and analysis of robust adaptive internal model control schemes. The certainty
20
Adaptive internal model control 1.4
2
ÐÐy−system y -system output output ± ± ±ym−model ym-model output r(t)=1 r
t 1
1.2
ÐÐy−system y -system output output ± ± ±ym−model ym-model output output r(t)=sin(0.2t) r
t sin
0:2t
1.5
1 1
output value
output value
0.5 0.8
0.6
0
−0.5 0.4 −1
0.2
0
−1.5
0
50 100 time t (second)
150
−2
0
50 100 time t (second)
150
time t (second)
time t (second)
Figure 1.4 MRAC IMC simulation
1.4
y−system output
ÐÐ y-system r(t)=1 output r
t 1
1.2
1
output value
0.8
0.6
0.4
0.2
0
−0.2 0
50
100
150 time t (second)
time t (second)
Figure 1.5 2 IMC simulation
200
250
300
Adaptive Control Systems 2
1.5
y−system output ÐÐ y -system output r(t)=0.8sin(0.2t) r
t 0:8 sin
0:2t
1.5
1
1
0.5
output value
output value
output ÐÐy−system y -system output r(t)=1 r
t 1
0.5
0
0
−0.5
−0.5
−1
−1
21
0
50
100 150 time t (second)
200
time t (second)
250
−1.5
0
100
200
300
t (second) timetime t (second)
Figure 1.6 1 IMC simulation
equivalence approach of adaptive control was used to combine a robust adaptive law with robust internal model controller structures to obtain adaptive internal model control schemes with provable guarantees of robustness. Some speci®c adaptive IMC schemes that were considered here include those of the partial pole placement, model reference, 2 optimal and 1 optimal control types. A single analysis procedure encompassing all of these schemes was presented. We do believe that the results of this chapter complete our earlier work on adaptive IMC [4, 5] in the sense that a proper bridge has now been established between adaptive control theory and some of its industrial applications. It is our hope that both adaptive control theorists as well as industrial practitioners will derive some bene®t by traversing this bridge.
References [1] Morari, M. and Za®riou, E. (1989). 0 . Prentice-Hall, Englewood Clis, NJ.
22
Adaptive internal model control
[2] Takamatsu, T., Shioya, S. and Okada, Y. (1985). `Adaptive Internal Model Control and its Application to a Batch Polymerization Reactor', 1.0 . / 0 + 0( , Frankfurt am Main. [3] Soper, R. A., Mellichamp, D. A. and Seborg, D. E. (1993). `An Adaptive Nonlinear Control Strategy for Photolithography', 2 .2 0 0 +2 [4] Datta, A. and Ochoa, J. (1996). `Adaptive Internal Model Control: Design and Stability Analysis', . , Vol. 32, No. 2, 261±266, Feb. [5] Datta, A. and Ochoa, J. (1998). `Adaptive Internal Model Control: 2 Optimization for Stable Plants', . . Vol. 34, No. 1, 75±82, Jan. [6] Youla, D. C., Jabr, H. A. and Bongiorno, J. J. (1976). `Modern Wiener-Hopf Design of Optimal Controllers ± Part II: The Multivariable Case', 1333 2 . 2 0 2, Vol. AC-21, 319±338. [7] Zames, G. and Francis, B. A. (1983). `Feedback, Minimax Sensitivity, and Optimal Robustness', 1333 2 . 2 0 2, Vol. AC-28, 585±601. [8] Desoer, C. A. and Vidyasagar, M. (1975). ! 1 45 , Academic Press, New York. [9] Ioannou, P. A. and Datta, A. (1991). `Robust Adaptive Control: A Uni®ed Approach', 2 + ( 1333, Vol. 79, No. 12, 1736±1768, Dec. [10] Kharitonov, V. L. (1978). `Asymptotic Stability of an Equilibrium Position of a Family of Systems of Linear Dierential Equations', 67 8 9 / , Vol. 14, 2086±2088. [11] Middleton, R. H., Goodwin, G. C., Hill, D. J. and Mayne, D. Q. (1988) `Design Issues in Adaptive Control', 1333 2 . 2 0 2, Vol. AC-33, 50±58. [12] Ioannou, P. A. and Sun, J. (1996). . / 0 , Prentice Hall, Englewood Clis, NJ. [13] Tsakalis, K. S. (1992). `Robustness of Model Reference Adaptive Controllers: An Input-Output Approach', 1333 2 . 2 0 2, Vol. AC-37, 556±565. [14] Ioannou, P. A. and Tsakalis, K. S. (1986). `A Robust Direct Adaptive Controller,' 1333 2 . 2 0 2, Vol. AC-31, 1033±1043, Nov.
2
An algorithm for robust adaptive control with less prior knowledge G. Feng, Y. A. Jiang and R. Zmood
Abstract A new robust discrete-time singularity free direct adaptive control scheme is proposed with respect to a class of modelling uncertainties in this chapter. Two key features of this scheme are that a relative dead zone is used but no knowledge of the parameters of the upper bounding function on the class of modelling uncertainties is required, and no knowledge of the lower bound on the leading coecient of the parameter vector is required to ensure the control law singularity free. Global stability and convergence results of the scheme are provided.
2.1
Introduction
Since it was shown (e.g. [1], [2]) that unmodelled dynamics or even a small bounded disturbance could cause most of the adaptive control algorithms to go unstable, much eort has been devoted to developing robust adaptive control algorithms to account for the system uncertainties. As a consequence, a number of adaptive control algorithms have been developed, for example, see [3] and references therein. Among those algorithms are simple projection (e.g. [4], [5]), normalization (e.g. [6], [7]), dead zone (e.g. [8±12]), adaptive law modi®cation (e.g. [13], [14]), -modi®cation (e.g. [15], [16]), as well as persistent excitation (e.g. [17], [18]). In the case of the dead zone based methods, a ®xed dead zone can be used [6±8] in the presence of only bounded disturbance, which turns o the algorithm when the identi®cation error is smaller than a certain threshold. In
24
order to choose an appropriate size of the dead zone, an upper bound on the disturbance must be known. When unmodelled dynamics are present, a relative dead zone modi®cation should be employed [11], [12]. Here the knowledge of the parameters of bounding function on the unmodelled dynamics and bounded disturbances is required. However, such knowledge, especially knowledge of the nonconservative upper bound or the parameters of the upper bounding function, can be hardly obtained in practice. Therefore, the robust adaptive control algorithm which does not rely on such knowledge is in demand but remains absent in the literature. One may argue that the robustness of the adaptive control algorithms can be achieved with only simple projection techniques in parameter estimation [4], [5]. However, it should be noted that using the robust adaptive control algorithms such as the dead zone, the robustness of the resulting adaptive control systems will be improved in the sense that the tolerable unmodelled dynamics can be enlarged [19]. Therefore, discussion of the robust adaptive control approaches such as those based on the dead zone technique is still of interest and the topic of this chapter. Another potential problem associated with adaptive control is its control law singularity. The estimated plant model could be in such a form that the polezero cancellations occur or the leading coecient of the estimated parameter vector is zero. In such cases, the control law becomes singular and thus cannot be implemented. In order to secure the adaptive control law singularity free, various approaches have been developed. These approaches can be classi®ed into two categories. One relies on persistent excitation. The other depends on modi®cations of the parameter estimation schemes. In the latter case, the most popular method is to hypothesize the existence of a known convex region in which no pole-zero cancellations occur and then to develop a convergent adaptive control scheme by constraining the parameter estimates inside this region (e.g. [11], [20±22]) for pole placement design; or to hypothesize the existence of a known lower bound on the leading coecient of the parameter vector and then to use an ad hoc projection procedure to secure the estimated leading coecient bounded away from zero and thus achieve the convergence and stability of the direct adaptive control system. However, such methods suer the problem of requirement for signi®cant a priori knowledge about the plant. Recently, another approach has been developed which also modi®es the parameter estimation algorithm. This approach is to re-express the plant model in a special input±output representation and then use a correction procedure in the estimation algorithm to secure the controllability and observability of the estimated model of the system [23±24]. They also addressed the robustness problem of such algorithms with respect to bounded disturbance [25] using the dead zone technique. They did not address the robustness problem with respect
25
to unmodelled dynamics. Moreover, those algorithms also suer the same problem as the usual dead zone based robust adaptive control algorithms. That is, they still require the knowledge of the upper bound on the disturbance or the parameters of the upper bounding function on the unmodelled dynamics and disturbances. In this chapter, a new robust direct adaptive control algorithm will be proposed which does use dead zone but does not require the knowledge of the parameters of the upper bounding functions on the unmodelled dynamics and the disturbance. It has also been shown that our algorithm can be combined with the parameter estimate correction procedure, which was originated in [24] to ensure the control law singularity free, so that the least a priori information is required on the plant. The chapter is organized as follows. The problem is formulated in Section 2.2. Ordinary discrete time direct adaptive control algorithm with dead zone is reviewed in Section 2.3. Our main results, a new robust direct adaptive control algorithm and its improved version with control law singularity free are presented in Section 2.4 and Section 2.5 respectively. Section 2.6 presents one simulation examples to illustrate the proposed adaptive control algorithms, which is followed by some concluding remarks in Section 2.7.
2.2
Problem formulation
Consider a discrete time single input single output plant yt
zd Bz1 ut vt Az1
2:1
where yt and ut are plant output and input respectively, vt represents the class of unmodelled dynamics and bounded disturbances and d is the time delay. Az1 and Bz1 are polynomials in z1 , written as Az1 1 a1 z1 . . . an zn Bz1 b1 z1 b2 z2 . . . bm zm Specify a reference model as Ez1 y t zd Rz1 rt where Ez1 is a strictly stable monic polynomial written as Ez1 1 e1 z1 . . . enzn
2:2
26
Then, there exist unique polynomials Fz1 and Gz1 written as Fz1 1 f1 z1 . . . fd1 zd1 Gz1 g0 g1 z1 . . . gn1 zn1
such that
Ez1 Fz1 Az1 zd Gz1
2:3
Using equation (2.3), it can be shown that the plant equation (2.1) can be rewritten as yt d z1 yt z1 ut T t t d
2:4
where yt d Ez1 yt d t d Fz1 Az1 vt d tT ut; ut 1; . . . ; ut m d 1; yt; yt 1; . . . ; yt n 1 : ut; tT T 1 ; . . . nmd : 1 ; T z1 Gz1 z1 Fz1 Bz1 We make the following standard assumptions [26], [18]. (A1) The time delay d and the plant order n are known. (A2) The plant is minimum phase. For the modelling uncertainties, we assume only: (A3) There exists a function [11] t such that
t 2 t where t satis®es
t "1 sup
x
2 "2 0t
for some unknown constants "1 > 0; "2 > 0; and x(t) is de®ned as xt yt 1; . . . ; yt n; ut 1; . . . ; ut m dT For the usual direct adaptive control, in order to facilitate the implementation of projection procedure to secure the control law singularity free, the following assumption is required.
27
(A4) There is a known constant 1m satisfying
1m 1
and
1m 1 > 0
For the usual relative dead zone based direct adaptive control algorithm, another assumption is needed as follows: (A5) The constants "1 and "2 in (A3) are known a priori. Remark 2.1 It should be noted that the assumptions (A4) and (A5) will not be required in our new adaptive control algorithm to be developed in the next few sections. It is believed that the elimination of assumptions (A4) and (A5) will improve the applicability of the adaptive control systems.
2.3
Ordinary direct adaptive control with dead zone
^ denote the estimate of the unknown parameter for the plant model Let t (2.4). De®ning the estimation error as ^ 1 et yt T t dt and a dead zone function as e g f g; e 0 eg
if e > g if e g
2:5
0 0
where the term at is a dead zone, which is de®ned as follows: 0 if et 2 t at otherwise f 1=2 t1=2 ; et=et with 0 < < 1,
2:7
2:8
0 ; 0 > 1, and proj is the projection operator [26] 1
28
such that ^1 t
^1 t if ^1 t sgn 1m 1m
1m
otherwise
2:9
It has been shown that the above parameter estimation algorithm has the following properties: ^ is bounded (i) t (ii) ^1 t 1m and (iii) (iv)
1 1 1 1 ^ ^ t t 1 1
f 1=2 t1=2 ; et2
1 t dT Pt 1t d f 1=2 t1=2 ; "t2 1 t dT Pt 1t d
where
l2
l2
^ dT t d "t yt t
2:10
The direct adaptive control law can be written as ut where
rf t ^ tT t ^1 t
rf t Rz1 rt
2:11 2:12
1
with rt a reference input and Rz is speci®ed by the reference model equation (2.2). Then with the parameter estimation properties (i)±(iv), the global stability and convergence results of the adaptive control systems can be established as in [26], [11], which are summarized in the following theorem. Theorem 3.1 The direct adaptive control system satisfying assumptions (A1)± (A5) with adaptive controller described in equations (2.7)±(2.9) and (2.11) is globally stable in the sense that all the signals in the loop remain bounded. However, as discussed in the ®rst section, the requirement for knowledge of the parameters of the upper bounding function on the unmodelled dynamics and bounded disturbances is very restrictive. In the next section, we are attempting to propose a new approach to get rid of such a requirement.
2.4
New robust direct adaptive control
Here we develop a new robust adaptive control algorithm which does not need such knowledge. That is, we drop the assumption (A5).
29
The key idea is to use an adaptation law to update those parameters. The new parameter estimation algorithm is the same as equations (2.7±2.9) but with a dierent dead zone at, where 0 if et 2 ^
t Qt at 2:13 1=2 1=2 f ^
t Qt ; et=et otherwise and Qt
41 1 t dT Pt 1t d
sup
x
2 0t
1
T
sup
x
2 0t
1
2:14 And ^tis calculated by sup
x
2 ^ T 0t
^t Ct 1 ^ Ct ^ 1 Ct
2:15
at 21 1 t dT Pt 1t d
sup
x
2
0t
1
> 0 2:16
where
^ T ^ Ct "1
"^2
with zero initial condition. It should be noted that "^1 and "^2 will be always positive and non decreasing. As shown in [23], the projection operation does not alter the convergence properties of the original parameter estimation algorithms. Therefore, in the following analysis, the projection operation will be neglected. The properties of the above modi®ed least squares parameter estimator with a relative dead zone are summarized in the following lemma. Lemma 4.1 The least squares algorithm with equations (2.7), (2.9), (2.13)± (2.16) applied to any system has the following properties irrespective of the control law: ^ is bounded (i) t ^ is bounded and non-decreasing, and thus "^1 converges to a constant, (ii) Ct say "1 1 1 1 (iii) ^1 t 1m and 1 1 ^ ^ t 1 t ^ t ^ 1
l2 (iv)
t
30
(v) f~t2 : Proof
f t2 1 t dT Pt 1t d
:
f 1=2 ^
t Qt1=2 ; et2 1 t dT Pt 1t d
l2
De®ne a Lyapunov function candidate ~ T Pt 11 t ~ Ct ~ 1 ~ 1T 1 Ct Vt 1 12 t
~ t ^ , Ct ~ 1 Ct ^ 1 "1 where t becomes Vt 1 Vt
2:17
"2 . Then, its dierence
at T
1 t d Pt 1t d
1 t dT Pt 1t d 2 2 t et 1 1 att dT Pt 1t d sup
x
2 ~ T atCt
0t 1 1 t dT Pt 1t d 1 at2
41 2 1 t dT Pt 1t d2 T sup
x
2 sup
x
2 0t 0t 1
1
at
T
1 t d Pt 1t d
1 t dT Pt 1t d 2 2 et t 1 1 att dT Pt 1t d
at T
~ T Ct
1 1 t d Pt 1t d sup
x
2 Qt 0t 1
at 1 2 2 t et 1 t dT Pt 1t d 1
at 1 1 t dT Pt 1t d
^
t t Qt
31
at T
1 t d Pt 1t d
1 1 2 2 t et ^
t t Qt 1 1
at 1 2 ^
t Qt et 1 1 t dT Pt 1t d
at 1 1 2 2 et et 1 1 t dT Pt 1t d
at 1 2 2 et et 0 1 t dT Pt 1t d
0 1 at et2 0 1 t dT Pt 1t d
0 1 f t2 0 1 t dT Pt 1t d
2:18
where the fact that atet2 f tet f t2 has been used. Therefore, following the same arguments in [20], [21], [23], the results in Lemma 4.1 are thus proved. If using the same adaptive control law as in equation (2.11), then with the parameter estimation properties (i)±(v) in Lemma 4.1, the global stability and convergence results of the new adaptive control system can be established as in [26], [11] as long as the estimated "1 is small enough, which are summarized in the following theorem. Theorem 4.1 The direct adaptive control system satisfying assumptions (A1)± (A4) with the adaptive controller described in equations (2.7), (2.9), (2.13)± (2.16) and (2.11) is globally stable in the sense that all the signals in the loop remain bounded. In this approach, we have eliminated the requirement for the knowledge of the parameters of the upper bounding function on the modelling uncertainties. But the requirement for the knowledge of the lower bound on the leading coecient of the parameter vector, i.e. assumption (A4) is still there. In the next section, the technique of the parameter correction procedure will be combined with the algorithm developed in this section to ensure the least prior knowledge on the plant. That is, only assumptions (A1)±(A3) are needed.
32
2.5
Robust adaptive control with least prior knowledge
The following modi®ed least squares algorithm will be used for robust parameter estimation: ^ t ^ 1 at t
Pt 1t d 1 t dT Pt 1t d
Pt Pt 1 at P1 k0 I;
et
Pt 1t dt dT Pt 1
2:19
1 t dT Pt 1t d
k0 > 0
and the parameter estimate is then corrected [24] as where
t ^ Pt t t
2:20
1T t d et yt t
2:21
the vector t is described in Figure 2.1 t t
pt pt
t 0 "pt
" ^1 t
pt 2 ^1 t
Figure 2.1
where pt is the ®rst column of the covariance matrix Pt, and the term at is now de®ned as follows: 0 if et 2 ^
t Qt at : 1=2 1=2
t Qt ; et= et otherwise f ^ 20 ; 0 > 1, and 1 Qt t 1T Pt 1t d2
with 0 < < 1,
21 1 t dT Pt 1t d
sup
x
2 0t
1
T
sup
x
2 0t
1
And ^tis calculated by sup
x
2 ^ T 0t
^t Ct 1 ^ Ct ^ 1 Ct
33
2:22
at 1 1 t dT Pt 1t d
sup
x
2
0t
1
;
> 0 2:23
where ^ T ^ Ct "1
"^2
with zero initial condition. It should be noted that "^1 and "^2 will be always positive and non-decreasing. Remark 5.1 The prediction error et is used in the modi®ed least squares algorithm to ensure that the estimator property (iii) in the following lemma can be established. The properties of the above modi®ed least squares parameter estimator are summarized in the following lemma. Lemma 5.1 If the plant satis®es the assumptions (A1)±(A3), the least squares algorithm (2.19)±(2.23) has the following properties: ^ is bounded, and
t ^ t ^ 1
l2 . (i) t ^ is bounded and non-decreasing, thus converges (ii) Ct f 1=2 ^
t Qt1=2 ; et2
(iii)
1 t dT Pt 1t d (iv)
pt
^1 t > bmin where
l2
bmin
1
max1;
with de®ned such that ^ Pt t 1" bmin (v) 1 t > 3" is bounded, and
t t 1
l2 (vi) t Proof
De®ne a Lyapunov function candidate ~ T Pt1 t ~ Ct ~ 1T 1 Ct ~ 1 Vt 1 12 t
2:24
34
~ t ^ , Ct ~ 1 Ct ^ 1 "1 where t
"2 T . Noting that
1T t d et yt t ^ 1t d t 1T Pt 1t d yt t et t 1T Pt 1t d
2:25
Then, the dierence of the Lyapunov function candidate becomes Vt 1 Vt
at T
1 t d Pt 1t d
1 t dT Pt 1t d 1 1 att dT Pt 1t d t t 1T Pt 1t d2 et2
~ T atCt 1 1 t dT Pt 1t d at2
sup
x
2
0t
sup
x
2
0t
1 at
T
sup
x
2
0t
1
1
T
1 t d Pt 1t d
1 t t 1T Pt 1t d2 et2 1 2at^
t t
1 1 t dT Pt 1t d at2 1 2 1 t dT Pt 1t d2
sup
x
2
0t
1
T
sup
x
2
0t
1
1 2 1 t dT Pt 1t d2
35
at T
1 t d Pt 1t d 2 2 t2 t 1T Pt 1t d2 et2 1 1
2at^
t t 1 1 t dT Pt 1t d at2 1 2 1 t dT Pt 1t d2 at
sup
x
2
0t
1
T
sup
x
2
0t
1
T
1 t d Pt 1t d
2 2 t2 et2 ^
t t Qt 1 1
at 2 2 ^
t Qt e t 1 1 t dT Pt 1t d
at 2 1 2 2 et et 1 1 t dT Pt 1t d
at 1 2 2 e t e t 0 1 t dT Pt 1t d
0 1 at et2 T 0 1 t d Pt 1t d
0 1 f 1=2 ^
t Qt1=2 ; et2 0 1 t dT Pt 1t d
2:26
where the fact that atet2 f tet f t2 has been used. Therefore, following the same arguments in [26], [11], [21], the results (i)±(iii) in Lemma 5.1 are thus proved. The properties (iv)±(vi) in the lemma can also be obtained directly from the results in [23]. If using the same adaptive control law as in equation (2.11), then with the parameter estimation properties (i)±(vi), the global stability and convergence results of the new adaptive control system can be established as in [26], [11] as long as the estimated "1 is small enough, which are summarized in the following theorem.
36
Theorem 5.1 The direct adaptive control system satisfying assumptions (A1)± (A3) with the adaptive controller described in equations (2.19)±(2.23) and (2.11) is globally stable in the sense that all the signals in the loop remain bounded.
2.6
Simulation example
In this section, one numerical example is presented to demonstrate the performance of the proposed algorithm. A fourth order plant is given by the transfer function as Gs Gn s Gu s with 5s 2 Gn s ss 1 as a nominal part, and 229 Gu s 2 s 30s 229 as the unmodelled dynamics. With the sampling period T 0:1 second, we have the following corresponding discrete-time model Gq1
0:09784q1 0:1206q2 0:1414q3 0:01878q4 1 2:3422q1 1:0788q2 0:4906q3 0:04505q4
The reference model is chosen as Gm s
1 0:2s 1
whose corresponding discrete-time model is Gm q1
0:3935q1 1 0:6065q1
^ 0:6 0 0 0T . If no dead zone is used, We have chosen k0 1, and 0 the simulation results are divergent. If using the algorithm developed in this chapter with 105 , the simulation results are shown as in Figure 2.2, where (a) represents the system output yt and reference model output y t, (b) is the control signal ut, (c) is the estimated parameter ^1 , and (d) denotes the estimated bounding parameters "^1 and "^2 . In order to demonstrate the eect of the update rate parameter , the following simulation with 1:4 105 was also conducted. The result is shown in Figure 2.3. The steady state values of the several important parameters and the tracking error in both cases are summarized in Table 2.1.
Control
Output and reference
Time in seconds
Estimated theta1
Est. bounding parameters
Time in seconds
Time in seconds
Time in seconds
Control
Output and reference
Figure 2.2 10
5
Time in seconds
Estimated theta1
Est. bounding parameters
Time in seconds
Time in seconds
Time in seconds
Figure 2.3 1:4 10 5
37
38
Table 2.1
"^1 "^2 ^1
y y
105
1:4 105
1.1898 10 3 0.3509 10 3 0.5649 0.0179
1.4682 10 3 0.4467 10 3 0.5833 0.07587
It can be observed from the above simulation results that the algorithm developed in this chapter can guarantee the stability of the adaptive system in the presence of the modelling uncertainties, and the smaller tracking error could be achieved with smaller update rate parameter . Most importantly, the knowledge of the parameters "1 and "2 of the upper bounding function and the knowledge of the leading coecient of the parameter vector 1 are not required a priori.
2.7
Conclusions
In this chapter, a new robust discrete-time direct adaptive control algorithm is proposed with respect to a class of unmodelled dynamics and bounded disturbances. Dead zone is indeed used but no knowledge of the parameters of the upper bounding function on the unmodelled dynamics and disturbances is required a priori. Another feature of the algorithm is that a correction procedure is employed in the least squares estimation algorithm so that no knowledge of the lower bound on the leading coecient of the plant numerator polynomial is required to achieve the singularity free adaptive control law. The global stability and convergence results of the algorithm are established.
References [1] Rohrs, C., Valavani, L., Athans, M. and Stein, G. (1985). `Robustness of Adaptive Control Algorithms in the Presence of Unmodelled Dynamics', IEEE Trans. Automat. Contr., Vol. AC-30, 881±889. [2] Egardt, B. (1979). Stability of Adaptive Controllers, Lecture Notes in Control and Information Sciences, New York, Springer Verlag. [3] Ortega, R. and Tang, Y. (1989). `Robustness of Adaptive Controllers ± A Survey', Automatica, Vol. 25, 651±677. [4] Ydstie, B. E. (1989). `Stability of Discrete MRAC-revisited', Systems and Control Letters, Vol. 13, 429±438.
39
[5] Naik, S. M., Kumar, P. R., Ydstie, B. E. (1992). `Robust Continuous-time Adaptive Control by Parameter Projection', IEEE Trans. Automat. Contr., Vol. AC-37, No. 2, 182±197. [6] Praly, L. (1983). `Robustness of Model Reference Adaptive Control', Proc. 3rd Yale Workshop on Application of Adaptive System Theory, New Haven, Connecticut. [7] Praly, L. (1987). `Unmodelled Dynamics and Robustness of Adaptive Controllers', presented at the Workshop on Linear Robust and Adaptive Control, Oaxaca, Mexico. [8] Petersen, B. B. and Narendra, K. S. (1982). `Bounded Error Adaptive Control', IEEE Trans. Automat. Contr., Vol. AC-27, 1161±1168. [9] Samson, C. (1983). `Stability Analysis of Adaptively Controlled System Subject to Bounded Disturbances', Automatica, Vol. 19, 81±86. [10] Egardt, B. (1980). `Global Stability of Adaptive Control Systems with Disturbances', Proc. JACC, San Francisco, CA. [11] Middleton, R. H., Goodwin, G. C., Hill, D. J. and Mayne, D. Q. (1988). `Design Issues in Adaptive Control', IEEE Trans. Automat. Contr., Vol. AC-33, 50±58. [12] Kreisselmeier, G. and Anderson, B. D. O. (1986). `Robust Model Reference Adaptive Control', IEEE Trans. Automat. Contr., Vol. AC-31, 127±133. [13] Kreisselmeier, G. and Narendra, K. S. (1982). `Stable Model Reference Adaptive Control in the Presence of Bounded Disturbances', IEEE Trans. Automat. Contr., Vol. AC-27, 1169±1175. [14] Iounnou, P. A. (1984). `Robust Adaptive Control', Proc. Amer. Contr. Conf., San Diego, CA. [15] Ioannou, P. and Kokotovic, P. V. (1984). `Robust Redesign of Adaptive Control', IEEE Trans. Automat. Contr., Vol. AC-29, 202±211. [16] Iounnou, P. A. (1986). `Robust Adaptive Controller with Zero Residual Tracking Error', IEEE Trans. Automat. Contr., Vol. AC-31, 773±776. [17] Anderson, B. D. O. (1981). `Exponential Convergence and Persistent Excitation', Proc. 20th IEEE Conf. Decision Contr., San Diego, CA. [18] Narendra, K. S. and Annaswamy, A. M. (1989). Stable Adaptive Systems, PrenticeHall, NJ. [19] Feng, G. and Palaniswami, M. (1994). `Robust Direct Adaptive Controllers with a New Normalization Technique', IEEE Trans. Automat. Contr., Vol. 39, 2330±2334. [20] Goodwin, G. C. and Sin, K. S. (1981) `Adaptive Control of Nonminimum Phase Systems', IEEE Trans. Automat. Contr., Vol. AC-26, 478±483. [21] Feng, G. and Palaniswami, M. (1992). `A Stable Implementation of the Internal Model Principle', IEEE Trans. Automat. Contr., Vol. AC-37, 1220±1225. [22] Feng, G., Palaniswami, M. and Zhu, Y. (1992). `Stability of Rate Constrained Robust Pole Placement Adaptive Control Systems', Systems and Control Letters, Vol. 18, 99±107. [23] Lazono-Leal, R. and Goodwin, G. C. (1985). `A Globally Convergent Adaptive Pole Placement Algorithm without a Persistency of Excitation Requirement', IEEE Trans. Automat. Contr., Vol. AC-30, 795±799.
40
[24] Lazono-Leal, R., Dion, J. and Dugard, L. (1993). `Singularity Free Adaptive Pole Placement Using Periodic Controllers', IEEE Trans. Automat. Contr., Vol. AC-38, 104±108. [25] Lazono-Leal, R. and Collado, J. (1989). `Adaptive Control for Systems with Bounded Disturbances', IEEE Trans. Automat. Contr., Vol. AC-34, 225±228. [26] Goodwin, G. C. and Sin, K. S. (1984). Adaptive Filtering, Prediction and Control, Prentice-Hall, NJ. [27] Lazono-Leal, R. (1989). `Robust Adaptive Regulation without Persistent Excitation', IEEE Trans. Automat. Contr., Vol. AC-34, 1260±1267.
3
Adaptive variable structure control C.-J. Chien and L.-C. Fu
3.1
Introduction
In the past two decades, model reference adaptive control (MRAC) using only input/output measurements has evolved as one of the most soundly developed adaptive control techniques. Not only has the stability property been rigorously established [17], [19] but also the robustness issue due to unmodelled dynamics and input/output disturbance has been successfully solved [15], [18]. However, several limitations on MRAC remain to be relaxed, especially the problem of unpredictable transient response and tracking performance which has recently become one of the challenging research topics in the ®eld of MRAC. A considerable amount of eort has been made to improve these schemes to obtain better control eects [6], [9], [11], [22]. One eort out of several is to try to incorporate the variable structure design (VSD) [9], [11] concept into the traditional model reference adaptive controller structure. Notably, Hsu and Costa [11] have ®rst successfully proposed a plausible scheme in this line, which was then followed by a series of more general results [12], [13], [14]. Aside from those, Fu [9], [10] has taken up a dierent approach in placing the variable structure design in the overall resulting adaptive controller. An ospring of the work [9] and part of the work [12] include various versions of results respectively applied to SISO [20], [23], MIMO [2], [5], time-varying [4], decentralized [24] and ane nonlinear [3] systems. It is well known that a main diculty for the design of the variable structure MRAC system is the so-called general case when relative degree of the plant is greater than one. In this chapter, we present a new algorithm to solve the variable structure model reference adaptive control for a single input single output system with unmodelled dynamics and output disturbances. The design concept will be ®rst introduced for relative degree-one plants and then be
42
Adaptive variable structure control
extended to the general case. Compared with the previous works, which used adaptive variable structure design or traditional robust adaptive approaches for the MRAC problem, this algorithm has the following special features: (1) This control algorithm successfully applies the variable structure adaptive controller for the general case under robustness consideration. (2) The control strategy using the concept of `average control' rather than that of `equivalent control' is thoroughly analysed. (3) A systematic design approach is proposed and a new adaptation mechanism is developed so that the prior upper bounds on some appropriately de®ned but unavailable system parameters are not needed. It is shown that without any persistent excitation the global stability and robustness with asymptotic tracking performance can be guaranteed. The output tracking error can be driven to zero for relative degree-one plants and to a small residual set (whose size depends on the level of magnitude of some design parameter) for plants with any higher relative degree. Both results are achieved even when the unmodelled dynamic and output disturbance are present. (4) If the aforementioned bounds on the system parameters are available by some means before controller design, then with a suitable choice of initial control parameters, the output tracking error can even be driven to zero in ®nite time for relative degree-one plants and to a small residual set exponentially for plants with any higher relative degree. It is noted that these bounds are usually assumed to be known before the construction of the variable structure controller or the robust adaptation law. In order to make a comparison between the proposed adaptive variable structure scheme and the traditional approaches, some computer simulations are made to illustrate the dierences of the tracking performance. The simulations will clearly demonstrate the excellent transient responses as well as tracking performance, which are almost never possible to achieve when traditional MRAC schemes are employed [19]. The theoretical framework in this chapter is developed based on Filippov's solution concept for a dierential equation with discontinuous right-hand side [8]. In the subsequent discussions, the following notations will be used: (1) Psu
t: denotes the ®ltered version of u
t with any proper or strictly proper transfer function P
s. (2) j j: denotes the absolute value of any scalar or the Euclidean norm of any vector or matrix. (3) k
t k1 supt j
j: denotes the truncated L1 norm of the argument function or vector. (4) kP
sk1 : denotes the H1 norm of the transfer function P
s. The chapter is organized as follows. In Section 3.2, we give the plant
Adaptive Control Systems
43
description, control objective and then derive the MRAC based error model. In Section 3.3, the adaptive variable structure controller for relative degree-one plants is proposed with stability and performance analysis. The extension to plants with relative degree greater than one is presented in Section 3.4. Section 3.5 gives simulation results to demonstrate the eectiveness of the adaptive variable structure controller. Finally, a conclusion is made in Section 3.6.
3.2
Problem formulation
3.2.1 Plant description and control objective In this chapter, we consider the following SISO linear time-invariant plant described by the equation: yp t P
s 1 Pu
s up
t do
t
3:1 where up
t and yp
t are plant input and plant output respectively, Pu
s is the multiplicative unmodelled dynamics with some 2 R , and do is the output disturbance. Here, P
s represents the strictly proper rational transfer function of the nominal plant which is described by P
s kp
np
s dp
s
3:2
where np
s and dp
s are some monic coprime polynomials and kp is the high frequency gain. Now suppose that the plant (3.1) is not precisely known but some prior knowledge about the transfer function may be available. The control objective is to design an adaptive variable structure control scheme such that the output yp
t of the plant will track the output ym
t of a linear time-invariant reference model described by ym
t M
srm
t km
nm
s rm
t dm
s
3:3
where M
s is a stable transfer function and rm
t is a uniformly bounded reference input. In order to achieve such an objective, we need some assumptions on the modelled part of the plant and the reference model as well as the unmodelled part of the plant. These assumptions are made in the following. For the modelled part of the plant and reference model: (A1) All the coecients of np
s and dp
s are unknown a priori, but the order of P
s and its relative degree are known to be n and , respectively. Without loss of generality, we will assume that the order of M
s and its relative degree are also n and , respectively.
44
Adaptive variable structure control
(A2) The value of high frequency gain kp is unknown, but its sign should be known. Without loss of generality, we will assume kp , and hence km , are positive. (A3) Ps is minimum phase, i.e. all its zeros lie in the open left half complex plane. For the unmodelled part of the plant: (A4) The unmodelled dynamics Pu s k1 is a strictly proper and stable transfer matrix such that jDj < a1 ; k
Pu
s k1 s D
s a2 k1 < a1 , for some constants a1 ; a2 > 0, where D lims!1 Pu
ss and kX
sk1 supw2R jX
jwj [15]. (A5) The output disturbance is dierentiable and the upper bounds on d jdo
tj; do
t exist. dt
Remark . Minimum-phase assumption (A3) on the nominal plant P
s is to guarantee the internal stability since the model reference control involves the cancellation of the plant zeros. However, as commented by [15], this assumption does not imply that the overall plant (3.1) possesses the minimum-phase property. . The latter part of assumption (A4) is simply to emphasize the fact that Pu
s are uncorrelated with in any case [16]. The reasons for assumption (A5) will be clear in the proof of Theorem 3.1 and that of Theorem 4.1.
3.2.2 MRAC based error model Since the plant parameters are assumed to be unknown, a basic strategy from the traditional MRAC [17], [19] is now used to construct the error model between yp
t and ym
t. Instead of applying the traditional MRAC technique, a new adaptive variable structure control will be given here in order to pursue better robustness and tracking performance. Let (3.1) be rewritten as 4
3:4 yp
t do
t P
s up Pu
sup
t P
sup u
t then from the traditional model reference control strategy [19], it can be shown that there exists 1 ; . . . ; 2n > 2 R2n such that if Db
s 1 ; 2 ; . . . ; n 1
a
s
s
Df
s n ; n1 ; . . . ; 2n 2
a
s 2n
s
1
where a
s 1; s; . . . ; sn 2 > and
s is an nth order monic Hurwitz
Adaptive Control Systems
45
polynomial, we have 1
Db
s
Df
sP
s 2n M 1
sP
s
3:5
Applying both sides of (3.5) on up u, we have up
t u
t Db
sup u
t Df
s yp do
t 2n M 1
s yp do
t
3:6
so that do
t M
s2n 1 up u
yp
t Since
Db
sup u
t Df
s yp
Db
sup u
Df
s yp
do
t
3:7
do
t 2n rm
t
a
s
s up u
t
a
s
> y d
t o
s p
yp
t do
t
rm
t
a
s
s up
t
a
s
> y
t
s p
yp
t
rm
t
4
> w
t
0
a
s do
t
s
D
s > u
t b
d
t
o 0
> wdo
t Db
s u
t
3:8
we have yp
t
do
t M
s2n 1 up M
s2n 1 up
> w > wdo
1
Db
s u 2n rm
t
> w > wdo
sup 2n rm
t
3:9
1 . . . n 1 sn 2 where
s
1 Db
sPu
s 1 Pu
s. If we de®ne
s the tracking error e0
t as yp
t ym
t, then the error model due to the unknown parameters, unmodelled dynamics and output disturbances can be
46
Adaptive variable structure control
readily found from (3.3) and (3.9) as follows: 1 > > e0 t M
s2n up w wdo
sup
t do
t
3:10
In the following sections, the new adaptive variable structure scheme is proposed for plants with arbitrary relative degree. However, the control structure is much simpler for relative degree-one plant, and hence in Section 3.3 we will ®rst give a discussion for this class of plants. Based on the analysis for relative degree-one plants, the general case can then be presented in a more straightforward manner in Section 3.4.
3.3
The case of relative degree one
When P
s is relative degree one, the reference model M
s can be chosen to be strictly positive real (SPR) (Narendra and Annaswamy, 1988). The error model (3.10) can now be rewritten as e0
t M
s2n 1 up > w > wdo 2n M 1
sdo
sup
t
3:11
>
>
2n M 1
sdo
and In the error model (3.11), the terms w; wdo
sup are the uncertainties due to the unknown plant parameters, output disturbance, and unmodelled dynamics, respectively. Let
Am ; Bm ; Cm be any minimal realization of M
s2n 1 which is SPR, then we can get the following state space representation of (3.11) as: _ Am e
tBm
up
t > w
t> wdo
t2n M 1
sdo
t
sup
t e
t e0
t Cm e
t
3:12
where the triplet
Am ; Bm ; Cm satis®es Pm Am A> m Pm P> m
2Qm ;
> Pm Bm Cm
3:13
Q> m
for some Pm > 0 and Qm > 0. The adaptive variable structure controller for relative degree-one plants is now summarized as follows: (1) De®ne the regressor signal > a
s a
s w
t up
t; yp
t; yp
t; rm
t w1
t; w2
t; . . . ; w2n
t>
s
s
3:14
Adaptive Control Systems
47
and construct the normalization signal mt [15] as the state of the following system: 1 _
3:15 mt 0 m
t 1
jup
tj 1 m
0 > 0 where 0 ; 1 > 0 and 0 2 < min
k1 ; k2 for some 2 > 0. The parameter k2 > 0 is selected such that the roots of
s k2 lie in the open left half complex plane, which is always achievable. (2) Design the control signal up
t as up
t
2n
sgn
e0 wj j
twj
t
sgn
e0 1
t
sgn
e0 2
tm
t
3:16
j1
1 0 sgn
e0 1
if e0 > 0 if e0 0 if e0 < 0
(3) The adaptation law for the control parameters is given as _j
t j je0
twj
tj;
j 1; . . . ; 2n
_1
t g1 je0
tj _2
t g2 je0
tjm
t
3:17
where j ; g1 ; g2 > 0 are the adaptation gains and j
0; 1
0; 2
0 > 0 (in general, as large as possible) j 1; . . . ; 2n. The design concept of the adaptive variable structure controller (3.15) and (3.16) is simply to construct some feedback signals to compensate for the uncertainties because of the following reasons: . By assumption (A5), it can be easily found that j> wdo
t 2n M
s 1 do
tj 1 for some 1 > 0. . With the construction of m, it can be shown [15] that
sup
t 2 m
t; 8t 0 and for some constant 2 > 0.
Now, we are ready to state our results concerning the properties of global stability, robust property, and tracking performance of our new adaptive variable structure scheme with relative degree-one system.
(Global Stability, Robustness and Asymptotic Zero Tracking Performance) Consider the system (3.1) satisfying assumptions (A1)±(A5) with relative degree being one. If the control input is designed as in (3.15), (3.16) and the adaptation law is chosen as in (3.17), then there exists > 0 such that for 2 0; all signals inside the closed loop system are bounded and the tracking error will converge to zero asymptotically.
48
Adaptive variable structure control
Consider the Lyapunov function Va 12 e> Pm e
2n 1
j 2 j j1
jj j2
2 1
j 2g j j1
j 2
where Pm satis®es (3.13). Then, the time derivative of Va along the trajectory (3.12) (3.17) will be V_ a
e> Qm e e0 up 2n 1
j
j1 j
e> Qm e
> w > wdo 2n M 1
sdo
sup
jj j_j
2 1 j1
2n
je0 wj j
j
gj
j
jj j
j _j je0 j
1
1
je0 j
2
2 m
j1
2n 1
j
j1 j
qm jej
2
jj j_j
2 1 j1
gj
j
j _j
for some constant qm > 0. This implies that e 2 L2 \ L1 and j ; j 1; . . . ; 2n; 1 ; 2 ; e0 2 L1 and, hence, all signals inside the closed loop system are bounded owing to Lemma A in the Appendix. On the other hand, it can be concluded that e_ 2 L1 by (3.12). Hence, e 2 L2 \ L1 and e_ 2 L1 readily imply that e and e0 will at least converge to zero asymptotically by Barbalat's lemma [19]. Q.E.D. In Theorem 3.1, suitable integral adaptation laws are given to compensate for the unavailable knowledge of the bounds on jj j and j . Theoretically, the adaptive variable structure controller will stabilize the closed loop system with guaranteed robustness and asymptotic zero tracking performance no matter what j
0's and j
0's are. However, according to the following Theorem 3.2, we will expect that positive and large values of j
0; j
0 should result in better transient response and tracking performance, especially when j
0 > jj j; j
0 > j .
(Finite-Time Zero Tracking Performance with High Gain Design) Consider the system set-up in Theorem 3.1. If j
0 jj j; j
0 j ; then the output tracking error will converge to zero in ®nite time with all signals inside the closed loop system remaining bounded. Proof
Consider the Lyapunov function Vb 12 e> Pm e where Pm satis®es
Adaptive Control Systems
49
(3.13). The time derivative of Vb along the trajectory (3.12) becomes V_ b
2n
e> Q m e
je0 wj j
j
jj j
je0 j
1
1
2 m
je0 j
2
j1
>
e Qm e
k3 Vb
for some k3 > 0 since j
t jj j; j
t j ; 8t 0. This implies that e approaches zero at least exponentially fast. Furthermore, by the fact that e0 e_0 e0 fCm Am e Cm Bm
up k4 je0 jjej
2n
je0 wj j
j
> w > wdo 2n M 1
sdo
sup g jj j
je0 j
1
1
je0 j
2
2 m
j1
k4 je0 jjej
je0 j
2n
jwj j
j
jj j
1
1
2
j1
2 m
where k4 jCm Am j, and that jej approaches zero at least exponentially fast, there exists a ®nite time T1 > 0 such that e0 e_0 k5 je0 j for all t > T1 and for some k5 > 0. This implies that the sliding surface e0 = 0 is guaranteed to be reached in some ®nite time T2 > T1 > 0. Q.E.D. Although theoretically only asymptotic zero tracking performance is achieved when the initial control parameters are arbitrarily chosen, it is encouraged to set the adaptation gains j and gj in (3.17) as large as possible. This is because the large adaptation gains will provide high adaptation speed and, hence, increase the control parameters to a suitable level of magnitude so as to achieve a satisfactory performance as quickly as possible. These expected results can be observed in the simulation examples.
3.4
The case of arbitrary relative degree
When the relative degree of (3.1) is greater than one, the controller design becomes more complicated than that given in Section 3.3. The main dierence between the controller design of a relative degree-one system and a system with relative degree greater than one can be described as follows. When (3.1) is relative degree-one, the reference model can be chosen to be strictly positive real (SPR) [19]. Moreover, the control structure and its subsequent analysis of global stability, robustness and tracking performance are much simpler. On the contrary, when the relative degree of (3.1) is greater than one, the reference model M
s is no longer SPR so that the controller and the analysis technique in relative degree-one systems cannot be directly applied. In order to use the
50
Adaptive variable structure control
similar techniques given in Section 3.3, the adaptive variable structure controller is now designed systematically as follows: (1) Choose an operator L1 s l1
s . . . l 1
s
s 1 . . .
s 1 such that M
sL1
s is SPR and denote Li
s li
s . . . l 1
s; i 2; . . . ; 1; L
s 1. (2) De®ne augmented signal 1 ya
t M
sL1
s u1 up
t L1
s and auxiliary errors ea1
t e0
t ya
t
3:18
ea2
t
1 u2
t l1
s
1 u1
t F
s
3:19
ea3
t
1 u3
t l2
s
1 u2
t F
s
3:20
.. . ea
t
1 l 1
s
up
t
1 u 1
t F
s
3:21
1 ui
t is the average control of ui
t with F
s
s 12 , F
s being small enough. In fact, F
s can be any Hurwitz polynomial in s 1 is referred to with degree at least two and F
0 1. In the literature, F
s as an averaging ®lter, which is obviously a low-pass ®lter whose bandwidth can be arbitrarily enlarged as ! 0. In other words, if is smaller and 1 smaller, the ®lter is ¯atter and ¯atter. F
s where
(3) Design the control signals up ; ui , and the bounding function m as follows: u1
t
2n
sgn
ea1 j j
tj
t
sgn
ea1 1
t
sgn
ea1 2
tm
t
j1
ui
t
li 1
s ui 1
t ; i 2; . . . ; sgn
eai F
s
up
t u
t
3:22
3:23
3:24
Adaptive Control Systems
51
with > 0 and t
1 1 1 w
t w
t l1
s l 1
s L1
s
The bounding function m
t is designed as the state of the system _ m
t
0 m
t 1
jup
tj 1; m
0 >
1 0
3:25
with 0 ; 1 > 0 and 0 2 < min
k1 ; k2 ; 1 ; . . . ; 1 for some 2 > 0. (4) Finally, the adaptation law for the control parameters j ; j 1; . . . ; 2n and 1 ; 2 are given as follows: _j
t j jea1
tj
tj;
j 1; . . . ; 2n
3:26
_1
t g1 jea1
tj
3:27
_2
t g2 jea1
tjm
t
3:28
with j
0 > 0; j
0 > 0 and j > 0; gj > 0. In the following discussions, the construction of feedback signals
t; m
t and the controller (3.22) (3.23) will be clear. In order to analyse the proposed adaptive variable structure controller, we ®rst rewrite the error model (3.10) as follows: e0
t M
sup
2n 1 > w 2n 1 > wdo 2n 1
sup
2n 1
1up
t do
t 1 1 up 2n 1 > 2n > wdo 2n M 1
sdo M
sL1
s L1
s L1
s 1 2n
sup
1 2n up
t
3:29 L1
s Now, according to the design of the above auxiliary error (3.18) and error model (3.29), we can readily ®nd that ea1 always satis®es 1 ea1
t M
sL1
s u1 2n 1 > 2n > wdo 2n M 1
sdo L1
s 2n 1
sup
1 2n up
t
3:30 L1
s It is noted that the auxiliary error ea1 is now explicitly expressed as the output term of a linear system with SPR transfer function M
sL1
s driven by some uncertain signals due to unknown parameters, output disturbances, unmodelled dynamics and unknown high frequency gain sign.
Adaptive variable structure control
52
The construction of the adaptive variable structure controller (3.22) is now clear since the following facts hold: 2n 1 > wdo 2n M 1
sdo
t is uniformly bounded due to (A5), L1
s we have 1 2n > 1
3:31 w M
sd do o
t 1 2n L
s
. Since
1
for some 1 . . With the design of the bounding function m
t (3.25), it can be shown that 1 2n L
s
sup
1 1
2n up
for some 2 > 0.
t 2 m
t
3:32
The results described in Remark 4.1 show that the similar techniques for the controller design of a relative degree-one system can now be used for auxiliary error ea1 . But what happens to the other auxiliary errors ea2 ; . . . ; ea , especially the real output error e0 as concerned? In Theorem 4.1, we summarize the main results of the systematically designed adaptive variable structure controller for plants with relative degree greater than one.
(Global Stability, Robustness and Asymptotic Tracking Performance) Consider the nonlinear time-varying system (3.1) with relative degree > 1 satisfying (A1)±(A5). If the controller is designed as in (3.18)± (3.25) and parameter update laws are chosen as in (3.26)±(3.28), then there exists > 0 and > 0 such that for all 2
0; and 2
0; , the following facts hold: (i) (ii) (iii) (iv)
all signals inside the closed-loop system remain uniformly bounded; the auxiliary error ea1 converges to zero asymptotically; the auxiliary errors eai ; i 2; . . . ; , converge to zero at some ®nite time; the output tracking error e0 will converge to a residual set asymptotically whose size is a class K function of the design parameter .
Proof
The proof consists of three parts.
Part I
Prove the boundedness of eai and 1 ; . . . ; 2n ; 1 ; 2 .
Step 1
First, consider the auxiliary error ea1 which satis®es (3.30). Since
Adaptive Control Systems
53
MsL1 s is SPR, we have the following realization of (3.20) 1 e_1 A1 e1 B1 u1 2n 1 > 2n > wdo 2n M 1
sdo L1
s
1 2n
sup
1 2n up L1
s ea1 C1 e1
3:33
2Q1 ; P1 B1 C1> for some P1 P> with P1 A1 A> 1 P1 1 > 0 and Q1 > Q1 > 0. Given a Lyapunov function as follows: 2 2 2n j 1 1 1 > V 1 2 e 1 P 1 e1 j
j j 2
3:34 2 2g j j 2n j1 j1
it can be shown by using (3.32) and (3.31) that 1 > _ V1 e1 Q1 e1 ea1 u1 2n 1 > 2n > wdo 2n M 1
sdo L1
s
2n 1
sup
1 2n up L1
s 2n 2 j 1 1 j _j
j j _j
g 2n j1 j j1 j 2n j je j e> Q e jea1 j
1 1 jea1 j
2 2 m a1 j j 1 1 1 2n j1 2 2n 1 1 j j _j
j j _j
g j j 2n j1 j1
e> 1 Q 1 e1
q1 je1 j2
for some q1 > 0 if the controller in (3.22) and update laws in (3.26)±(3.28) are given. This implies that e1 ; 1 ; . . . ; 2n ; 1 ; 2 2 L1 and ea1 2 L2 \ L1 . Step 2
From (3.19)±(3.21), we can ®nd that ea2 ; . . . ; ea satisfy e_a2
1 ea2 u2
l1
s u1 F
s
.. . e_a
1 ea u
l 1
s u 1 F
s
54
Adaptive variable structure control
Now by the following facts that for i 2 . . . ; : d 1 2 e eai e_ai dt 2 ai eai i 1 eai ui or
li 1
s ui 1 F
s
li 1
s 2 ui 1 i 1 eai eai sgn
eai F
s d jeai j dt
i 1 jeai j
li 1
s ui 1 F
s
3:35
when jeai j 6 0. This implies that eai will converge to zero after some ®nite time T > 0. Part II Prove the boundedness of all signals inside the closed loop system. De®ne eai M
sLi 1
seai ; i 2; . . . ; and Ea ea1 ea2 ea which is uniformly bounded due to the boundedness of eai . Then, from (3.18)±(3.21), we can derive that 1 up Ea e0 M
sL1
s u1 L1
s 1 1 u1 u2 M
sL1
s l1
s F
s 1 1 u2 M
sL2
s u3 l2
s F
s .. . M
sL 1
s e0 1 4
e0 R
1
l 1
s
up
1 u 1 F
s
1 1 1 u2 u 1 M
sL1
s u1 F
s l1
s l1
s . . . l 2
s
3:36
Now, since k
ui t k1 k6 k
e0 t k1 k6 ; i 1; . . . ; 1 for some k6 > 0 by Lemma A in the appendix, it can be easily found that
1 u1 1 u2 u 1 k7 k
e0 t k1 k7 l1
s l1
s . . . l 2
s t 1
Adaptive Control Systems
55
1 k1 2 and for some k7 > 0. Furthermore, since the H1 norm of k 1s
1 F
s ksM
sL1
sk1 k8 for some k8 > 0, we can conclude that
1 1 k
Rt k1 1 sM
sL1
s
k7 k
e0 t k1 k7 s F
s 1 1
k9 k
e0 t k1 k9
for some k9 > 0. Now from (3.36) we have k
e0 t k1 k
Ea t k1 k
Rt k1 k
Ea t k1
k9 k
e0 t k1 k9 which implies that there exists a > 0 such that 1 2
0; : k
Ea t k1 k9 k
e0 t k1 1 k9
k9 > 0 and for all
3:37
Combining Lemma A and (3.37), we readily conclude that all signals inside the closed loop system remain uniformly bounded. Part III: Investigate the tracking performance of ea1 and e0 . Since all signals inside the closed loop system are uniformly bounded, we have ea1 2 L2 \ L1 ; e_a1 2 L1 Hence, by Barbalat's lemma, ea1 approaches zero asymptotically and Ea ea1 ea2 ea also approaches zero asymptotically. Now, from the fact of (3.37) and Ea approaching zero, it is clear that e0 will converge to a small residual set whose size depends on the design parameter . Q.E.D. As discussed in Theorem 3.2, if the initial choices of control parameters j j
0; j
0 satisfy the high gain conditions j
0 and j
0 j , then, 2n
by using the same argument given in the proof of Theorem 3.2, we can guarantee the exponential convergent behaviour and ®nite-time tracking performance of all the auxiliary errors eai . Since eai reaches zero in some ®nite time and Ea ea1 ea2 ea , it can be concluded that Ea converges to zero exponentially and e0 converges to a small residual set whose size depends on the design parameter . We now summarize the results in the following Theorem 4.2.
(Exponential Tracking Performance with High Gain Design) Consider the system set-up in Theorem 4.1. If the initial value of control j parameters satisfy the high gain conditions j
0 and j
0 j , then 2n
there exists a and such that for all 2
0; and 2
0; , the following facts hold: (i) all signals inside the closed loop system remain bounded;
56
Adaptive variable structure control
(ii) the auxiliary errors eai i 1 . . . ; , converge to zero in ®nite time; (iii) the output tracking errors e0 will converge to a residual set exponentially whose size depends on the design parameter . It is well known that the chattering behaviour will be observed in the input channel due to variable structure control, which causes the implementation problem in practical design. A remedy to the undesirable phenomenon is to introduce the boundary layer concept. Take the case of relative degree one, for example, the practical redesign of the proposed adaptive variable structure controller by using boundary layer design is now stated as follows:
2n
e0 1
t
e0 2
tm
t
3:38
e0 wj j
twj
t up
t j1
sgn
e0 if je0 j > "
e0 e0 if je0 j " "
for some small > 0. Note that
e0 is now a continuous function. However, one can expect that the boundary layer design will result in bounded tracking error, i.e. e0 cannot be guaranteed to converge to zero. This causes the parameter drift in parameter adaptation law. Hence, a leakage term is added into the adaptation law as follows: _j
t j je0
twj
tj _1
t g1 je0
tj
j
t; j 1; . . . ; 2n
1
t
_2
t g2 je0
tjm
t
2
t
3:39
for some > 0.
3.5
Computer simulations
The adaptive variable structure scheme is now applied to the following unstable plant with unmodelled dynamics and output disturbances:
8 1 yp
t 3 1 0:01 up
t 0:05 sin
5t s s2 s 2 s 10 Since the nominal plant is relative degree three, we choose the following steps to design the adaptive variable structure controller:
Adaptive Control Systems
57
. reference model and reference input:
Ms
8
s 23 2 if t < 5 rm
t 2 if 5 t < 10
. design parameters:
L1
s l1
sl2
s; l1
s s 1; l2
s s 2
s
s 12
2 1 s1 F
s 60 . augmented signal and auxiliary errors: ya
t M
sL1
s u1
1 up
t L1
s
ea1
t e0
t ya
t ea2
t
1 u2
t l1
s
1 u1
t F
s
ea3
t
1 u3
t l2
s
1 u2
t F
s
. controller:
u1
t
6 j1
ui
t
sgn
ea1 j j
tj
t
sgn
ea1 1
t
li 1
s ui 1
t 1 ; sgn
eai F
s
sgn
ea1 2
tm
t
i 2; 3
up
t u
t _ m
t
m
t 0:005
jup
tj 1;
m
0 0:2
. adaptation law:
_j
t j jea1
tj
tj;
j 1; . . . ; 6
_1
t g1 jea1
tj _2
t g2 jea1
tjm
t Three simulation cases are studied extensively in this example in order to verify
58
Adaptive variable structure control
Figure 3.1 yp
; ym
Figure 3.2 yp
; ym
, time (sec)
, time (sec)
all the theoretical results and corresponding claims. All the cases will assume that there are initial output error yp 0 ym
0 4. (1) In the ®rst case, we arbitrarily choose the initial control parameters as j
0 0:1;
j 1; . . . ; 6
j
0 0:1;
j 1; 2
and set all the adaptation gains j gj 0:1. As shown in Figure 3.1 (the time trajectories of yp and ym ), the global stability, robustness, and asymptotic tracking performance are achieved. (2) In the second case, we want to demonstrate the eectiveness of a proper choice of j
0 and j
0 and repeat the previous simulation case by increasing the values of the controller parameters to be j
0 1;
j 1; . . . ; 6
j
0 1;
j 1; 2
The better transient and tracking performance between yp and ym can now be observed in Figure 3.2.
Adaptive Control Systems
Figure 3.3 yp
; ym ;
59
, time (sec)
(3) As commented in Remark 3.2, if there is no easy way to estimate the suitable initial control parameters j 0 and j 0 like those in the second simulation case, it is suggested to use large adaptation gains in order to increase the adaptation rate of control parameters such that the nice transient and tracking performance as described in case 2 can be retained to some extent. Hence, in this case, we use the initial control parameters as in case 1 but set all the adaptation gains to j gj 1. The expected results are now shown in Figure 3.3, where rapid increase of control parameters do lead to satisfactory transient and tracking performance.
3.6
Conclusion
In this chapter, a new adaptive variable structure scheme is proposed for model reference adaptive control problems for plants with unmodelled dynamic and output disturbance. The main contribution of the chapter is the complete version of adaptive variable structure design for solving the robustness and performance of the traditional MRAC problem with arbitrary relative degree. A detailed analysis of the closed-loop stability and tracking performance is given. It is shown that without any persistent excitation the output tracking error can be driven to zero for relative degree-one plants and driven to a small residual set asymptotically for plants with any higher relative degree. Furthermore, under suitable choice of initial conditions on control parameters, the tracking performance can be improved, which are hardly achievable by the traditional MRAC schemes, especially for plants with uncertainties.
Adaptive variable structure control
60
Appendix Consider the controller design in Theorem 3.1 or 4.1. If the control parameters j t; j 1; . . . ; 2n; 1
t and 2
t are uniformly bounded 8t, then there exists > 0 such that up
t satis®es k
up t k1 k
e0 t k1
A:1
with some positive constant > 0. Consider the plant (3.1) which is rewritten as follows:
Proof
y
t
do
t P
s
1 Pu
sup
t
A:2
Let f
s be the Hurwitz polynomial with degree n such that f
sP
s is proper, and hence, f 1
sP 1
s is proper stable since P
s is minimum phase by assumption (A3). Then y
t
do
t P
sf
sf
1
s
1 Pu
sup
t
A:3
which implies that f
1
sP 1
s y
do
t
f
1
sPu
sup
t f
1
4
sup
t u
t
A:4
Since f 1
sP 1
s and f 1
sPu
s are proper or strictly proper stable, we can ®nd by small gain theorem [7] that there exists > 0 such that k
u t k1 k
yp t k1 k
e0 t k1
A:5
for some suitably de®ned > 0 and for all 2 0; . Now if we can show that k
up t k1 k
u t k1
A:6
for some > 0, then (A.1) is achieved. By using Lemma 2.8 in [19], the key point to show the boundedness between up and u in (A.6) is the growing behaviour of signal up . The above statement can be stated more precisely as follows: if up satis®es the following requirement jup
t1 j cjup
t1 Tj
A:7
where t1 and t1 T are the time instants de®ned as t1 ; t1 T ft j jup j k
up t k1 g
A:8
and c is a constant 2
0; 1, then up will be bounded by u , i.e. (A.6) is achieved. Now in order to establish (A.7) and (A.8), let
Ap ; Bp ; Cp and
; B be the a
s respectively. Also state space realizations of P
s
1 Pu
s and
s
Adaptive Control Systems
61
> > > de®ne S x> p w1 w2 m . Then, using the augmented system Bp up 0 xp x_ p Ap 0 0
w_ 1 0 0 0
w1 Bup
w_ 2 BCp 0 Bdo 0 w2
1 jup j 1 0 0 0
0 m_ m
Since do is uniformly bounded, we can easily show according to the control design (3.16) or (3.24) that there exists such that _ k
S k jSj t 1 This means that S is regular [21] so that xp w1 w2 m yp and up will grow at most exponentially fast (if unbounded), which in turn guarantees (A.7) and (A.8) by Lemma 2.8 in [19]. This completes our proof. Q.E.D.
References [1] Chien, C. J. and Fu, L. C., (1992) `A New Approach to Model Reference Control for a Class of Arbitrarily Fast Time-varying Unknown Plants', Automatica, Vol. 28, No. 2, 437±440. [2] Chien, C. J. and Fu, L. C., (1992) `A New Robust Model Reference Control with Improved Performance for a Class of Multivariable Unknown plants', Int. J. of Adaptive Control and Signal Processing, Vol. 6, 69±93. [3] Chien, C. J. and Fu, L. C., (1993) `An Adaptive Variable Structure Control for a Class of Nonlinear Systems', Syst. Contr. Lett., Vol. 21, No. 1, 49±57. [4] Chien, C. J. and Fu, L. C., (1994) `An Adaptive Variable Structure Control of Fast Time-varying Plants', Control Theory and Advanced Technology, Vol. 10, No. 4, part I, 593±620. [5] Chien, C. J., Sun, K. S., Wu, A. C. and Fu, L. C., (1996) `A Robust MRAC Using Variable Structure Design for Multivariable Plants', Automatica, Vol. 32, No. 6, 833±848. [6] Datta, A. and Ioannou, P. A., (1991) `Performance Improvement versus Robust Stability in Model Reference Adaptive Control', Proc. CDC, 748±753. [7] Desoer, C. A. and Vidyasagar, M., (1975) Feedback Systems: Input±Output Properties, Academic Press, NY. [8] Filippov, A. F., (1964) `Dierential Equations with Discontinuous Right-hand Side', Amer. Math. Soc. Transl., Vol. 42, 199±231. [9] Fu, L. C., (1991) `A Robust Model Reference Adaptive Control Using Variable Structure Adaptation for a Class of Plants', Int. J. Control, Vol. 53, 1359±1375. [10] Fu, L. C., (1992) `A New Robust Model Reference Adaptive Control Using Variable Structure Design for Plants with Relative Degree Two', Automatica, Vol. 28, No. 5, 911±926. [11] Hsu, L. and Costa, R. R., (1989) `Variable Structure Model Reference Adaptive Control Using Only Input and Output Measurement: Part 1', Int. J. Control, Vol. 49, 339±419.
62
Adaptive variable structure control
[12] Hsu, L., (1990) `Variable Structure Model-Reference Adaptive Control (VSMRAC) Using Only Input Output Measurements: the General Case', IEEE Trans. Automatic Control, Vol. 35, 1238±1243. [13] Hsu, L. and Lizarralde, F., (1992) `Redesign and Stability Analysis of I/O VSMRAC Systems', Proc. American Control Conference, 2725±2729. [14] Hsu, L., de Araujo A. D. and Costa, R. R., (1994) `Analysis and Design of I/O Based Variable Structure Adaptive Control', IEEE Trans. Automatic Control, Vol. 39, No. 1. [15] Ioannou, P. A. and Tsakalis, K. S., (1986) `A Robust Direct Adaptive Control', IEEE Trans. Automatic Control, Vol. 31, 1033±1043. [16] Ioannou, P. A. and Tsakalis, K. S., (1988) `The Class of Unmodeled Dynamics in Robust Adaptive Control', Proc. of American Control Conference, 337±342. [17] Narendra, K. S. and Valavani, L., (1978) `Stable Adaptive Controller Design ± Direct Control', IEEE Trans. Automatic Control, Vol. 23, 570±583. [18] Narendra, K. S. and Annaswamy, A. M., (1987) `A New Adaptive Law for Robust Adaptation Without Persistent Excitation', IEEE Trans. Automatic Control, Vol. 32, 134±145. [19] Narendra, K. S. and Valavani, L., (1989) Stable Adaptive Systems, Prentice-Hall. [20] Narendra, K. S. and BoÆskovicÂ, J. D., (1992) `A Combined Direct, Indirect and Variable Structure Method for Robust Adaptive Control', IEEE Trans. Automatic Control, Vol. 37, 262±268. [21] Sastry, S. S. and Bodson, M., (1989) Adaptive Control: Stability, Convergence, and Robustness, Prentice-Hall, Englewood Clis, NJ. [22] Sun, J., (1991) `A Modi®ed Model Reference Adaptive Control Scheme for Improved Transient Performance', Proc. American Control Conference, 150±155. [23] Wu, A. C., Fu, L. C. and Hsu, C. F., (1992) `Robust MRAC for Plants with Arbitrary Relative Degree Using a Variable Structure Design', Proc. American Control Conference, 2735±2739. [24] Wu, A. C. and Fu, L. C., (1994) `New Decentralized MRAC Algorithms for LargeScale Uncertain Dynamic Systems', Proc. American Control Conference, 3389±3393.
4
Indirect adaptive periodic control D. Dimogianopoulos, R. Lozano and A. Ailon
Abstract In this chapter a new indirect adaptive control method is presented. This method is based on a lifted representation of the plant which can be stabilized using a simple performant periodic control scheme. The controller parameter's computation involves the inverse of the controllability/observability matrix. Potential singularities of this matrix are avoided by means of an appropriate estimates modi®cation. This estimates transformation is linked to the covariance matrix properties and hence it preserves the convergence properties of the original estimates. This modi®cation involves the singular value decomposition of the controllability/observability matrix's estimate. As compared to previous studies in the subject the controller proposed here does not require the frequent introduction of periodic n-length sequences of zero inputs. Therefore the new controller is such that the system is almost always operating in closed loop which should lead to better performance characteristics.
4.1
Introduction
The problem of adaptive control of possibly nonminimum phase systems has received several solutions over the past decade. These solutions can be divided into several dierent categories depending on the a priori knowledge on the plant, and on whether persistent excitation can be added into the system or not. Schemes based on persistent excitation were proposed in [1], [11] among others. This approach has been thoroughly studied and is based on the fact that convergence of the estimates to the true plant parameter values is guaranteed when the plant input and output are rich enough. Stability of the
64
Indirect adaptive periodic control
closed loop system follows from the unbiased convergence of the estimates. The external persistent excitation signal is then required to be always present, therefore the plant output cannot exactly converge to its desired value because of the external dither signal. This diculty has been removed in [2] using a selfexcitation technique. In this approach excitation is introduced periodically during some pre-speci®ed intervals as long as the plant state and/or output have not reached their desired values. Once the control objectives are accomplished the excitation is automatically removed. Stability of these type of schemes is guaranteed in spite of the fact that convergence of the parameter estimates to their true values is not assured. This technique has also been extended to the case of systems for which only an upper bound on the plant order is known in [12] for the discrete-time case and in [13] for the continuoustime case. Since adding extra perturbations into the system is not always feasible or desirable, other adaptive techniques not resorting to persistent exciting signals have been developed. Dierent strategies have been proposed depending on the available information on the system. When the parameters are known to belong to given intervals or convex sets inside the controllable regions in the parameter space, the schemes in [9] or [10] can be used, respectively. These controllers require a priori knowledge of such controllable regions. An alternative method proposes the use of a pre-speci®ed number of dierent controllers together with a switching strategy to commute among them (see [8]). This method oers an interesting solution for the cases when the number of possible controllers in the set is ®nite and available. In the general case the required number of controllers may be large so as to guarantee that the set contains a stabilizing controller. In general, very little is known about the structure of the admissible regions in the parameter space. This explains the diculties encountered in the search of adaptive controllers not relying on exciting signals and using the order of the plant as the only a priori information. In this line of research a dierent approach to avoid singularities in adaptive control has been proposed in [6] which only requires the order of the plant as available information. The method consists of an appropriate modi®cation to the parameter estimates so that, while retaining all their convergence properties, they are brought to the admissible area. The extension of this scheme to the stochastic case has been carried out in [7]. The extension of this technique to the minimum phase multivariable case can be found in [4]. This method does not require any a priori knowledge on the structure of the controllable region. It can also be viewed as the solution of a least-squares parameter estimation problem subject to the constraint that the estimates belong to the admissible area. The admissible area is de®ned here as those points in the parameter space whose corresponding Sylvester resultant matrix is nonsingular. The main drawback of the scheme presented in [6] is that the number of directions to be explored in
Adaptive Control Systems
65
the search for an appropriate modi®cation becomes very large as the order of the system increases. This is due essentially to the fact that the determinant of the Sylvester resultant matrix is a very complex function of the parameters. The method based on parameter modi®cation has also been previously used in [3] for a particular lifting plant representation. The plant description proposed in [3] has more parameters than the original plant, but has the very appealing feature of explicitly depending on the system's controllability matrix. Indeed, one of the matrix coecients in the new parametrization turns out to be precisely the controllability matrix times the observability matrix. Therefore, the estimates modi®cation can actually be computed straightforwardly without having to explore a large number of possible directions as is the case in [6]. It actually requires one on-line computation involving a polar decomposition of the estimate of the controllability matrix. However, no indication was given in [3] on how this computation can be eectively carried out. Recently, [14] presented an interesting direct adaptive control scheme for the same class of liftings proposed in [3]. As pointed out in [14] the polar decomposition can be written in terms of a singular value decomposition (SVD) which is more widely known. Methods to perform SVD are readily available. This puts the adaptive periodic controllers in [3] and [14] into a much better perspective. At this point it should be highlighted that even if persistent excitation is allowed into the system, the presented adaptive control schemes oer a better performance during the transient period by avoiding singularities. The adaptive controller proposed in [3] and [14] is based on a periodic controller. A dead-beat controller is used in one half of the cycle and the input is identically zero during the other half of the cycle. Therefore a weakness of this type of controllers is that the system is left in open loop half of the time. In this chapter we propose a solution to this problem. As compared to [3] and [14] the controller proposed here does not require the frequent introduction of periodic n-length sequences of zero inputs. The new control strategy is a periodic controller calculated every n-steps, n being the order of the system. For technical reasons we still have to introduce a periodic sequence of zero inputs but the periodicity can be arbitrarily large. As a result the new controller is such that the system is almost always operating in closed loop which should lead to better performance characteristics.
4.2
Problem formulation
Consider a discrete-time system, described by the following state-space representation: xt1 Axt but b0 v0t
4:1 yt cT xt v00t
66
Indirect adaptive periodic control
where x is the
n 1 state vector, u, y are respectively the plant's input and output and A b cT are matrices of appropriate dimensions. Signals v0 v00 can be identi®ed as perturbations belonging to a class, which will be discussed later. The state part of equation (4.1) is iterated n times:
42 xtn An xt An 1 b . . . b Ut An 1 b0 . . . b0 Vt0 n 1 Matrix A b . . . b is the system's (4.1) controllability matrix C
n n and similarly C An 1 b0 . . . b0 . Vectors Ut Vt0 are de®ned as: Ut u
t . . . u
t n 1 U 2 Rn Vt0 v0
t . . . v0
t n 1 V 0 2 Rn Working in a similar way as in (4.2), an alternative expression for xt can be obtained:
43 xt An xt n CUt n C0 Vt0 n introducing (4.3) into (4.2) becomes: xtn A2n xt
n
An CUt
n
An C0 Vt0
n
CUt C0 Vt0
44
The next step will be to express the system's output y for a time interval t t n 1, using (4.1): yt cT xt v00t
yt1 cT Axt cT but cT b0 v0t v00t1 .. . ytn
1
cT An 1 xt cT An 2 but . . . cT butn cT b0 v0tn
2
v00tn
2
cT An 2 b0 v0t . . .
1
which can be written as the following expression: where
Ytn Oxt GUt G0 Vt0 IVt00 O cT
n 1 T
c T A . . . cT A
is the system's observability matrix 0 0 cT b 0 nn G .. .. .. .. G2R . . . . cT A n 2 b
cT b
0
45
O 2 Rnn yt n Yt ... Yt 2 Rn
yt
46
47
1
and Vt00 is similarly de®ned as Vt0 but with v0 replaced by v00 . Identically, G0 is de®ned as G in (4.7) with b0 instead of b. From (4.5) we get: 0 00 Yt2n Oxtn GUtn G0 Vtn IVtn
48
Adaptive Control Systems
67
Since the state is not supposed to be measurable, we will obtain an expression that depends only on the system's output and input. Introducing (4.4) into the above it follows Yt2n OA2n xt
n
OAn CUt
n
OAn C0 Vt0
n
0 00 GUtn G0 Vtn IVtn
OCUt OC0 Vt0
49
whereas, by (4.5) Yt Oxt n GUt n G0 Vt0 n IVt00 n xt n O 1 Yt GUt n G0 Vt0 n IVt00 n
and
410
411
So ®nally, by substituting (4.11) into (4.9), one has: Yt2n DYt BUt B0 Ut where t 0 n 2n . . . and D OA2n O 1
N OAn C0 Vt0
n
B OC
B0 OAn C
n
GUtn Nt2n OA2n O 1 G
412
D B B0 2 Rnn
413
0 00 O 1 A2n O 1 G0 Vt0 n OC0 Vt0 G0 Vtn IVtn OA2n O 1 Vt00
n
414
The following control law is proposed for the plant: BUt
B0 Ut
n
415
GUtn Nt2n
416
DYt
The closed loop system becomes: Yt2n
whereas from (4.11) and (4.16), one has: 0 00 O 1 Nt2n IVtn xtn O 1 Nt2n G0 Vtn
1 Nt2n
417
1 where Nt2n is another noise term. Therefore the state xtn t 0 n 2n . . . is bounded by the (bounded) noise. In the ideal case xtn will be identically zero. From (4.3), it is clear that Ut is bounded and becomes zero in the ideal case.
2.1 Dead-beat control can induce large control inputs. For a smoother performance, one can use the following control strategy: BUt
DYt
B0 Ut
n
C Yt
GUt n
418
where C has all its eigenvalues inside the unit disc. The closed loop system is in this case:
419 Yt2n GUtn C Yt GUt n Nt2n In the ideal case the LHS of the above converges to zero and so does the state.
Indirect adaptive periodic control
68
4.3
Adaptive control scheme
In this section the adaptive control scheme based on the strategy previously developed is presented. For sake of simplicity an index k will be associated to t as follows: t kn k 0 1 2 . . . The plant representation (4.12) can be rewritten as: Yk1 k Nk1 Yk1 2 Rn
420 where:
. . . B .. B0 .. G .. D
2 Rn4n k Uk
1
Uk
2
Uk Yk 1 T
2 R4n
421
for k 0 1 2 . . .. The class of disturbances satis®es: kN
kk
422
The following a priori knowledge on the plant is required: Assumption 1: n and an upper bound in (4.22) is known Assumption 2: The existence of a lower bound b0 is required, so that b0 I BT B. The value of this bound, however, may not be known. The equations of the adaptive scheme are given in the order they appear in the control algorithm. Note that signals Yk k 1 are measurable.The prediction error: Ek Yk k 1 k 1 Ek 2 Rn
423 where the variables marked with a bar are the normalized versions of the original ones: Yk Yk 1 kk 1 k
k
1
k 1 1 kk 1 k
424
Least squares with dead zone: T w2k EkT Ek k 1 P2k 1 k
425
1
k 1 kk 1 k 0
k jwk j k
1 n
1 Tk 1 P 2k k 1 jwk j
426 if w 2k 2k
1 n 0 otherwise
427
Adaptive Control Systems
Pk Pk
1
k k
1
T k Pk 1 k 1 k 1 Pk T 1 k k 1 Pk 1 k
1 1
Pk 2 R4n4n
. . . T k Ek k 1 Pk Bk .. B0 k .. Gk .. Dk k 2 Rn4n
69
428
429
The forgetting factor can also be chosen to commute between 0 and 1 using the same scheduling variable as in [6]. Modi®cation of the parameters estimates 1
L
n 4n T Pk Lk Lk 0 L L2
3n 4n Bk Qk Sk
Sk 0
Qk QTk I
Q S 2 Rnn
k k Qk L1k LTk Bk B0 k Gk Dk
430
431
432
k
(4.30) is the so-called Cholesky decomposition which gives a triangular positive semi-de®nite matrix L and (4.31) is a polar decomposition. The ®rst one is often found in software packages used for control purposes, as it is a special case of the LU decomposition. The second one can be calculated in many ways. The way that is described here consists of calculating the singular value decomposition of B: Bk U k S k V k T . Then the matrices Qk and Sk can be calculated as follows: Qk U k V Tk
4.4
Sk V k S k V Tk
433
Adaptive control law
4.4.1 Properties of the identi®cation scheme In this section we will present the convergence properties of the parameter identi®cation algorithm proposed in the previous section. These properties will be essential in the stability of the adaptive control closed loop system. Let us de®ne the parametric distance at instant k as: k k and matrix Hk as: k Ek EkT Hk 2 R4n4n
434 Hk Hk 1 T 1 k Pk 1 k 1
The following properties hold:
k 1
70
Indirect adaptive periodic control
(1) The forgetting factor in (4.27) satis®es: T k k 1 P2k 1 k
0 k
1
435
(2) There is a positive de®nite function Vk de®ned as: Vk tr
Pk Hk that satis®es: Vk Vk 1 (3) The augmented error wk in (4.25) is bounded as follows:
w2k
k!1
2k
1 n 0
436
(4) The covariance matrix converges. 2 (5) The forgetting factor is such that: 1 k0 k k 1. (6) The modi®ed parameter estimate Bk satis®es the following: Bk BTk
b20 I 4V02
437
(7) k converges and: k Tk V02 I (8) The modi®ed prediction error E k de®ned as: k 1 k
E k Yk
1
E k 2 Rn
438
is bounded as follows: E kT E k 2w2k
439
Proofs of the above properties are given in [3].
4.4.2 Adaptive control strategy The expression (4.38) can be written as follows (see also (4.24)): Yk E k
Yk
k 1 k 1 k 2 k 1 k k 1 k k 2 k
1
k 2 1 k k 1 k
1
k 1 k
1
From (4.21) and (4.32): Yk E k
Bk 2 Uk
k
B0 k 2 Uk 3 Gk 2 Uk 1 1 kUk 2 Uk 3 Uk 1 Yk 2 k 2
k 1 k 1 kk 1 k 2
1
Dk 2 Yk
2
440
Adaptive Control Systems
71
Using (4.32), we obtain:
k
k 1 k 2 1 kk 1 k
1
k
k
k 1 k
2 2
k 1 k
1
1
k 2 LT k
k 1 LT k 1 k
2
Note that property (7) tells us that: kk 2 k 1 k ! 0 as k ! 1 Note also that the last term in (4.41) can be rewritten as:
1
441
k 1 LT k 1 k 2 LT k 1 k 1 k 2
LT k 2 LT k 1 k 1
k 2 k 1 LT k 1 k 1
442 T But L k 2 LT k 1 ! 0 because matrix Lk results from the Cholesky decomposition of the covariance matrix P in (4.30), which converges. On the other hand, using (4.25), term LT k 1 k 1 in (4.42) can be written as:
k 2 LT k
2
LT k 1 k 1 T
LT k 1 k 1 Tk 1 Pk 1 k 1 kTk 1 k kPk 1 k 1 k Tk 1 PT k 1 Pk 1 k 1 (using (4.24)) E T k Ek Tk 1 PT k 1 Pk 1 k 1
w2 k k ! 1
k
Using (4.41), (4.42), (4.43) and the facts that as and k
k 2 k 1 k 1, we can conclude that: where
k
k
2
k
k
k 1 k 1 k k k
2
k 1 k
1
2
k 1 k
k 2
LT k
2
2
k 1 ! 0
jwk j k k k LT k 1 k
443
1
444
445
and k k k ! 0 as k ! 1. From (4.40) and using (4.39) and the above we have: kYk
B k 2 Uk
B0 k 2 Uk 3 Gk 2 Uk 1 Dk 2 Yk 2 k 1 kUk 2 Uk 3 Uk 1 Yk 2 k E k k
k 2 k 1 k 1 k
p 2jwk j kk 2 k 1 k jwk j k k k
446 2
or equivalently kYk1
Bk 1 Uk 1 B0 k 1 Uk 2 Gk 1 Uk 1 kUk 1 Uk 2 Uk Yk 1 k
Dk 1 Yk 1 k
p 2jwk1 j kk
1
k k
jwk1 j k k k
447
72
Indirect adaptive periodic control
Using the certainty equivalence principle the adaptive control strategy could simply be obtained from (4.15) by replacing the true parameters by their estimates. However, for technical reasons that will appear clear later, we will have to introduce some sequences of inputs equal to zero. These sequences are introduced periodically after a time interval of arbitrary length L. We will assume that k k where k is a time instant large enough such that the RHS of (4.47) is small enough. This will be made clear in (4.50). Since some sequences of inputs equal to zero are periodically introduced let us assume that k k is the ®rst time instant after k that we apply the zero sequences de®ned below: Uk
2
0
Uk
1
Bk 11
Dk 1 Yk
Uk Uk1 Uk2 Uk3
UkL
2 1
0 1 B0 k1 Bk1 Dk1 Yk1 Uk 1 B0 k2 Bk2 Dk2 Yk2 Uk1 etc.
UkL UkL1 UkL2 UkL3
B0 k 1 Uk 2
UkL
1
448
0 1 BkL
1
DkL 1 YkL 0
1
B0 kL 1 UkL
2
1 B0 kL1 YkL1 BkL1 DkL1 UkL 1 B0 kL2 YkL2 BkL2 DkL2 UkL1 etc.
The convergence analysis will be carried out by contradiction. We will assume that the regressor vector k goes to in®nity and prove that this leads to a contradiction. Note that if kkk ! 1 then from (4.26) and (4.36) it follows that kwkk ! 0. Introducing the control strategy (4.48) into (4.47) we obtain Y k1 !0
449 1 kUk 1 0 0 Yk 1 k
Adaptive Control Systems
Note from (4.48) that Uk 1 Bk 11
Dk 1 Yk 1 . Therefore Uk by Yk 1 . Introducing this fact into (4.49) Y k1 1 k Yk 1 k
1
73
is bounded
450
or equivalently
kYk1 k
1 k Yk 1 k
451
where 0 is arbitrarily small. Using equation (4.11), (4.51) and the fact that Uk 0, one has kxkk kO 1 kkYk1 k k Yk 1 k
452
Hereafter to simplify notation, will generically represent bounded terms. may denote the bound for the system noises or the bound for dierent matrix norms. We will not give the exact expression for because it is irrelevant for the analysis purposes. Combining (4.52 ) and (4.3), we have x k Y k
453 k1 k 1 From (4.3) we get
0 C0 Vk1 CUk1 An xk1 xk2
is given in (4.48) where Uk1 0 1 U Y B
B U D k k1 k1 k1 k1 k1 1 Bk1 Yk1 Dk1 1 Bk1
1 k Yk 1 k
see
451 Dk1
454
455
Consequently one has (see (4.53), (4.54), (4.55)): x k Y k k 1 k2
456 We may proceed in the same way to obtain an expression for xk3 Consider the following expressions from (4.3), (4.48), (4.10) respectively: 0 C0 Vk2 CUk2 An xk2 xk3 0 1 Bk2 Uk2 Dk2 Bk2 Yk2 Uk1
0 00 G0 Vk1 IVk1 GUk1 Oxk1 Yk2
457
458
459
74
Indirect adaptive periodic control
From the above and using (4.53) and (4.55), we obtain: Y Ox GU k2 k1 k1 k Yk 1 k
From the above and (4.55)±(4.58) one has: x b3 k Y k k 1 k3 Introducing (4.48) and (4.59), one has x b3 k x k k 2 k3
460
461
462
It should be noted that b3 is a ®nite quantity which, for a given system, depends only on the length of the time interval k 2 k 3 . We may extend this procedure over the interval k 2 k 2 L . Then (4.62) becomes: x bL k x k
463 k 2L
k 2
where bL is a quantity similar to b3 in (4.62). Finally we may repeatthe same procedure for iL i 1 2 in order to obtain an expression for xk 2iL x bL x
464 k 2iL k 2
i 1L
Expressions (4.63) and (4.64) can be combined to obtain: i 1 i x
bL i 2 bL
465 k 2iL
bL kxk 2 k
bL
Factor
bL i 1
bL i 2 bL will be a sum of an in®nite number of terms. But as bL 1 thissum will be bounded. Hence we will have a bounded decreasing state xk 2iL as follows: i x 8i 1
466 k 2iL
bL kxk 2 k
From (4.66) we conclude that the state xk 2iL for i 1 1 is bounded in the limit by a term that depends on the noise. The states at other time instants inside the intervals k 2
i 1L k 2 iL can be proven to be bounded too, by using the procedure described in (4.56) up to (4.62). This clearly contradicts our assumption that diverges. Therefore all the signals remain bounded. Furthermore in view of the previous analysis, the smaller the noise upper bound, the smaller the state becomes in the limit.
4.5
Simulations
4.5.1 System without noise The performance of the proposed adaptive control scheme is illustrated in this section through a series of simulations. The following non-minimum phase
Adaptive Control Systems
system is considered: xt1
yt 1
3
1
2
0
xt
0xt
v00t
1 15
ut b0 v0t
75
467
The system's poles and zeroes are respectively (1, 2) and 1.5. In this way matrices B B0 G D are:
105 25 0 0 14 15 15 1 0 G D B B 225 45 1 0 30 31 25 15
468 The initial parameter estimates B0 B0 0 G0 D0 are chosen in such a way that the signs of the determinants of B0 and B are opposite. This will lead to important conclusions concerning the utility of the modi®cation of the parameters' estimates. So the initial parameter estimates are:
2 1 105 25 0 0 14 15
469 0 25 15 225 45 1 0 30 31 and initial conditions are: zero control input and state x0 100 100T for t 2 0 8. The parameters' modi®cation was used only when the determinant of the estimate of B was below a threshold of 0.1. It is clear that k in (4.32) can be multiplied by a scalar without losing the convergence properties. Such a coecient can be used to reduce the abrupt changes in the modi®ed parameter estimates and therefore improve the transient behaviour. In our case in (4.32) was multiplied by a coecient equal to 0.001. The initial value of the covariance matrix is chosen as P 220I8 , whereas there is no noise corrupting the system. On the other hand sequences of zero inputs have not been used in the simulations that follow. Figures that are referred hereafter are given in the appendix. Figures 4.1 and 4.2 show that modifying the parameter estimates helps to avoid unboundedness of the plant signals y
t and u
t. The results that are presented in Figure 4.2 clearly denote that at around k 22 or t 44 there is a value of det(B) smaller than 0.1 in the case of unmodi®ed parameters. The values of det(B) that follow are all very close to zero, a fact that leads to the behaviour shown in Figure 4.1. On the other hand, when the modi®cation of the paprameters estimates is applied, the critical value of det(B) at t 44 is avoided and the evolution of det(B) thereafter is clearly away from zero. 4.5.2 System with noise Simulation results are now presented in the case when the system is corrupted
u
t WITHOUT Mod
Indirect adaptive periodic control
u
t WITH Mod
y
t WITH Mod
y
t WITHOUT Mod
76
Figure 4.1 Up: WITHOUT modi®cation, Down: WITH modi®cation of parameter estimates
Det(B) WITHOUT Mod
1 0.5 0 −0.5 −1 0
5
10
15
20
25 k
30
35
40
45
50
30
35
40
45
50
k
Det(B) WITH Mod
1 0.5 0 −0.5 −1 0
5
10
15
20
25 k
k
Figure 4.2 Up: WITHOUT modi®cation, Down: WITH modi®cation of parameter estimates. Note t 2k
77
y
t WITH Mod
u
t WITH Mod
u
t WITHOUT Mod
y
t WITHOUT Mod
Adaptive Control Systems
Figure 4.3 Up: WITHOUT modi®cation, Down: WITH modi®cation of parameter estimates 1.5
Determinant of B
Determinant of B
1
0.5
0
−0.5
−1 0
10
20
30
40
50 k
60
70
80
90
100
k
Figure 4.4 e: WITHOUT modi®cation, : WITH modi®cation of parameter B estimate. Note: t 2k
78
Indirect adaptive periodic control
with noise. System (4.67) remains the same with the exception of b0 b 1 15T in (Figure 4.1) and of a sinusoidal noise described by: v0 t 00004 sin
2k 00003 cos
2k
470
v t 00004 sin
2k
471
00
00003 cos
2k
Initial conditions are the same as in the ideal case (i.e. zero control input and state x0 100 100T until t 8) and P 200I8 . Figure 4.3 shows the evolution of output and input signals and Figure 4.4 explains how values of determinant of B estimates close to zero can lead to huge output and input signals.
4.6
Conclusions
This chapter has presented an indirect adaptive periodic control scheme based on the lifted representation of the plant proposed in [3]. New arguments have been presented so that this technique appears in a better position as a solution to the long standing problem of singularities in adaptive control of nonminimum phase plants. Contrary to other techniques, the only a priori knowledge required on the plant is its order besides the standard controllability/observability assumption. As compared to previous studies in the subject [3], [14], the proposed controller does not require the frequent introduction of periodic n-length sequences of zero inputs. Furthermore simulations have shown that in practice there is no need to introduce sequences of inputs equal to zero. Therefore the new controller is such that the system always operates in closed loop which leads to better performance characteristics. Simulation results have also shown that the use of the proposed estimate modi®cation can signi®cantly reduce signal peaks during the transient period even in the case when signals y
t and u
t of the original unmodi®ed system do not become unbounded.
References [1] Kreisselmeier, G. and Smith, M. C., (1986) `Stable Adaptive Regulation of Arbitrary n-th Order Plants', IEEE Trans. Contr., Vol. AC-31, 299±305. [2] Kreisselmeier, G. (1989) `An Indirect Adaptive Controller with a Self-excitation Capability', IEEE Trans. Automat. Contr., Vol. AC-34, 524±528. [3] Lozano, R. (1989) `Robust Adaptive Regulation Without Persistent Excitation', IEEE Trans. Autom. Contr., Vol. AC-34, 1260±1267. [4] de Mathelin, M. and Bodson. M., (1994) `Multivariable Adaptive Control: Identi®able Parametrizations and Parameter Convergence', IEEE Transactions on Automatic Control., Vol. 39, No. 8, 1612±1617.
Adaptive Control Systems
79
[5] Weyer, E., Mareels, I. M. and Polderman, J. W., (1994) `Limitations of Robust Adaptive Pole Placement Control', IEEE Transactions on Automatic Control., Vol. 39, No 8, 1665±1671. [6] Lozano, R. and Zhao, X. H., (1994) `Adaptive Pole Placement Without Excitation Probing Signal', IEEE Transactions on Automatic Control., Vol. 39, No 1, 47±58. [7] Guo. L., (1996) `Self-convergence of Weighted Least Squares with Application to Stochastic Adaptive Control', IEEE Transactions on Automatic Control., Vol. 41, No 1, 79±89, January. [8] Morse, A.S. and Pait, F. M., (1994) `MIMO Design Models and Internal Regulators for Cyclicly Switched Parameter-Adaptive Control Systems', IEEE Trans. Autom. Contr., Vol. 39, No 9, 1809±1818, September. [9] Ossman K. A. and Kamen, E. W., (1987) `Adaptive Regulation of MIMO Linear Discrete-time Systems without Requiring a Persistent Excitation', IEEE Trans. Autom. Contr., Vol. AC-32, No 5, 397±404. [10] Middleton, R. H., Goodwin, G. C., Hill, D. J. and Mayne, D. Q., (1988) `Design Issues in Adaptive Control', IEEE Trans. Autom. Contr., Vol. 33, No 1, 50±58. [11] Giri, F., M'Saad, M., Dugard, L. and Dion, J. M., (1992) `Robust Adaptive Regulation with Minimal Prior Knowledge', IEEE Trans. Autom. Contr., Vol. AC37, 305±315. [12] Kreisselmeier, G., (1994) Parameter Adaptive Control: A Solution to the Overmodeling Problem', IEEE Trans. Autom. Contr., Vol. 39, No. 9, 1819±1826, September. [13] Kreisselmeier, G. and Lozano, R., (1996) `Adaptive Control of Continuous-time Overmodeled Plants', IEEE Trans. on Autom. Contr., Vol. 41, No. 12, 1779±1794, December. [14] Bayard, D., (1996) `Stable Direct Adaptive Periodic Control Using Only Plant Order Knowledge', International Journal of Adaptive Control and Signal Processing, Vol. 10, No. 6, 551±570, November±December. [15] Lancaster, P. and Tismenetsky M., (1985) The Theory of Matrices. 2nd ed. Academic Press, NY.
5
Adaptive stabilization of uncertain discrete-time systems via switching control: the method of localization P. V. Zhivoglyadov, R. H. Middleton and M. Fu
Abstract In this chapter a new systematic switching control approach to adaptive stabilization of uncertain discrete-time systems with parametric uncertainty is presented. Our approach is based on a localization method which is conceptually dierent from supervisory adaptive control schemes and other existing switching adaptive schemes. Our approach allows for slow parameter drifting, infrequent large parameter jumps and unknown bound on exogenous disturbance. The unique feature of localization based switching adaptive control distinguishing it from conventional adaptive switching schemes is its rapid model falsi®cation capabilities. In the LTI case this is manifested in the ability of the switching controller to quickly converge to a suitable stabilizing controller. We believe that the approach presented in this chapter is the ®rst design of a switching controller which is applicable to a wide class of linear time-invariant and time-varying systems and which exhibits good transient performance. The performance of the proposed switching controllers is illustrated by many simulation examples.
5.1
Introduction
Control design for both linear and nonlinear dynamic systems with unknown parameters has been extensively studied over the last three decades. Despite
Adaptive
81
signi®cant advances in adaptive and robust control in recent years, control of systems with large-size uncertainty remains a dicult task. Not only are the control problems complicated, so is the analysis of stability and performance. It is well known [12,14] that classical adaptive algorithms prior to 1980 were all based on the following set of standard assumptions or variations of them: (i) (ii) (iii) (iv)
An upper bound on the plant order is known. The plant is minimum phase. The sign of high frequency gain is known. The uncertain parameters are constant, and the closed loop system is free from measurement noise and input/output disturbances.
Classical adaptive algorithms are known to suer from various robustness problems [34]. A number of attempts have been made since 1980 to relax the assumptions above. A major breakthrough occurred in the mid-1980s [17, 21, 35] for adaptive control of LTV plants with suciently small in the mean parameter variations. Later attempts were made for a broader class of systems. Fast varying continuous-time plants were treated in [36], assuming knowledge of the structure of the parameter variations. By using the concept of polynomial dierential (integral) operators the problem of model reference adaptive control was dealt with in [32] for a certain class of continuous-time plants with fast time-varying parameters. An interesting approach based on some internal self-excitation mechanism was considered in [7] for a general class of LTV discrete-time systems. The global boundedness of the state was proved. However, it must be noted that the presence of such self-excitation signals in a closed loop system is often undesirable. In another research line, a number of switching control algorithms have been proposed recently by several authors [2, 6, 8, 20, 23, 24, 31], thus signi®cantly weakening the assumptions in (i)±(iv). Both continuous and discrete linear time-invariant systems were considered. Research in this direction was originated by the pioneering works of Nussbaum [31] and Martensson [20]. Nussbaum considered the problem of ®nding a smooth stabilizing controller
_ f gt; zt zt ut gyt; zt
5:1
for the one-dimensional system
_ axt qu
t xt y
t x
t
5:2
with both q 6 0 and a > 0 unknown. In [31] Nussbaum describes a whole family of controllers of the form (5.1) which achieve the desired stabilization of
82
Adaptive
the system (5.2). For example, it was shown that every solution xt; zt of x_ ax qx
z2 1 cos
z=2 exp z2
5:3 z_ x
z2 1 has the property that limt!1 x
t 0 and limt!1 z
t exists and is ®nite. We note that the structure of the adaptive controller is explicitly seen from (5.3). Another important result proved in [31] is that there exists no stabilizing controller for the plant (5.2) expressed in terms of polynomial or rational functions. A more general result was presented by Martensson [20]. In particular, it was shown in [20] that the only a priori information which is needed for adaptive stabilization of a minimal linear time-invariant plant is the order of a stabilizing controller. This assumption can even be removed if a slightly more complicated controller is used. Consider the following dynamic feedback problem. Given the plant x_ Ax Bu; x 2 n ; u 2 m ;
5:4 y Cx; y 2 r and the controller
z_ Fz Gy;
z 2 l
u Hz Ky
5:5
where m, r are known and ®xed, and n is allowed to be arbitrary. It is easy to see that this is equivalent to the static feedback problem ^x B^ ^u x^_ A^
5:6 y^ C^x^; u^ K^ y^ ^ B, ^ and K^ are matrices ^ C, where x^
xT zT T , u^
uT zT T , y^
yT zT T and A, of appropriate dimensions. Let the regulator be u^ g
h
kN
h
k^ y
5:7 2 2 _ k jj^ yjj jj^ ujj where N
h is an `almost periodic' dense function and h and g are continuous, scalar functions satisfying a set of four assumptions (see [20] for more details). Martensson's result reads: `Assume that l is known so that there exists a ®xed stabilizing controller of the form (5.5), and that the augmentation to the form (5.6) has been done. Then the controller (5.7) will stabilize the system in the sense that
x
t; z
t; k
t !
0; 0; k1 where k1 < 1'.
as t ! 1
5:8
Adaptive
83
One such set of functions given by Martensson is hk log k1=2 ; k 1; g
h
sin h1=2 1h1=2
5:9
Martensson's method is based on a `dense' search over the control parameter space, allows for no measurement noise, and guarantees only asymptotic stability rather than exponential stability. These weaknesses were overcome in [8] where a ®nite switching control method was proposed for LTI systems with uncertain parameters satisfying some mild compactness assumptions. Dierent modi®cations of Martensson's controller aimed at achieving Lyapunov stability, avoiding dense search procedures, as well as extending this approach to discrete-time systems have been reported recently (see, e.g., [2, 8, 19, 23]). However, the lack of exponential stability might result in poor transient performance as pointed out by many researchers (see, e.g., [8,19] for simulation examples). Below we present a simple example of a controller based on a dense search over the parameter space. This controller is a simpli®ed version of that presented in [19]. Example 1.1
The second order plant
x
t 1 a1 x
t a2 x
t
1 bu
t
t;
x; u 2 R
5:10
with a1;2 2 R; b 6 0 being arbitrary unknown constants and suptt0 j
tj < 1 has to be controlled by the switching controller u
t k
tx
t
5:11
where k
0 h
1 and k
t h
i, t 2
ti ; ti1 and h
i is a function dense in R de®ned so that it successively looks at each interval p; p; p 2 N and tries points 1=2p apart, namely, h
1 1
h
4
0:5
h
7 1:75
h
2 0:5
h
5
1
etc:
h
3 0
h
6 2
The system performance is monitored using a function
t M
ti 1
ti 1
t
ti 1
jx
ti 1 j
ti 1
5:12
For each i > 1 such that ti 1 6 1, the switching instant is de®ned as minft : t > ti 1 ; jx
tj > M
ti 1
ti 1
t ti jx
ti 1 j
ti 1 g if this exists ti 1 otherwise
5:13 where 0 < M
t, 0 <
t < 1 and 0 <
t are strictly positive increasing functions satisfying the following conditions limt!1 M
t 1,
84
Adaptive 9
1
x 10
2 controller gain
output
0.5 0 −0.5 −1 0
50 (a)
100
1 0 −1 −2 0
50 (b)
100
Figure 5.1
limt!1
t 1, limt!1
t 1. The behaviour of the closed loop system with a1 2:2, a2 0:3 and b 1 is illustrated in Figure 5.1(a)±(b). A dierent switching control approach, called hysteresis switching, was reported in a number of papers [22, 27, 37] in the context of adaptive control. In these papers, the hysteresis switching is used to swap between a number of `standard' adaptive controllers operating in regimes of the parameter space. The switching, in these cases, is used to avoid the `stabilizability' problem in adaptive controllers. Conventional switching control techniques are all based on some mechanism of an exhaustive search over the entire set of potential controllers (either a continuum set [20] or a ®nite set [8]). A major drawback is that the search may converge very slowly, resulting in excessive transients which renders the system `unstable' in a practical sense. This phenomenon can take place even if the closed loop system is exponentially stable. To alleviate this problem, several new switching control schemes have been proposed recently. The so-called supervisory control of LTI systems for adaptive set-point tracking is proposed by Morse [25, 26] to improve the transient response. A further extension of Morse's approach is given in [13]. A very similar, in spirit, supervisory control scheme for model reference adaptive control is analysed in [29]. The main idea of supervisory control is to orchestrate the process of switching into feedback controllers from a pre-computed ®nite (continuum) set of ®xed controllers based on certain on-line estimation. This represents a signi®cant departure from traditional estimator based tuning algorithms which usually employ recursive or dynamic parameter tuning schemes. This approach has apparently signi®cantly improved the quality of regulation, thus demonstrating that switching control if properly performed is no longer just a nice theoretical toy but a powerful tool for high performance control systems design. However, several issues still remain unresolved. For example:
Adaptive
85
(i) a ®nite convergence of switching is not guaranteed. This aspect is especially important in situations when convergence of switching is achievable. It seems intuitively that in adaptive control of a linear timeinvariant system it is desirable that the adaptive controller `converges' to a linear time-invariant controller; (ii) the analysis of the closed loop stability is quite complicated and often dependent on the system architecture. Without a simpler proof and better understanding of the `hidden' mechanisms of supervisory switching control its design will remain primarily a matter of trial and error. In this chapter, we present a new approach to switching adaptive control for uncertain discrete-time systems. This approach is based on a localization method, and is conceptually dierent from the supervisory control schemes and other switching schemes. The localization method was initially proposed by the authors for LTI systems [39]. This method has the unique feature of fast convergence for switching. That is, it can localize a suitable stabilizing controller very quickly, hence the name of localization. Later this method was extended to LTV plants in [40]. By utilizing the high speed of localization and the rate of admissible parameter variations exponential stability of the closed loop system was proved. The main contribution of this chapter is a uni®ed description of the method of localization. We show that this method is also easy to implement, has no bursting phenomenon, and can be modi®ed to work with or without a known bound on the exogenous disturbance. To highlight the principal dierences between the proposed framework and existing switching control schemes, in particular, supervisory switching control, we outline potential advantages of localization based switching control: (i) The switching controller is ®nitely convergent provided that the system is time invariant. Depending on how the switching controller is practically implemented the absence of this property could potentially have far reaching implications. (ii) Unlike conventional switching control based on an exhaustive search over the parameter space, the switching converges rapidly thus guaranteeing a high quality of regulation. (iii) The closed loop stability analysis is comparatively simple even in the case of linear time-varying plants. This is in sharp contrast to supervisory switching control where the stability analysis is quite complicated and depends on the system architecture. (iv) Localization based switching control is directly applicable to both linear time-invariant and time-varying systems. (v) The localization technique provides a clear understanding of the control mechanism which is important in applications. The rest of this chapter is organized as follows. Section 5.2 introduces the class
86
Adaptive
of LTI systems to be controlled and states the switching adaptive stabilization problem. Two dierent localization principles are studied in Sections 5.3 and 5.4. We also study a problem of optimal localization, which allows us to obtain guaranteed lower bounds on the number of controllers discarded at each switching instant and adaptive stabilization in the presence of unknown exogenous disturbance. Simulation examples are given in Section 5.5 to demonstrate the fast switching capability of the localization method. Conclusions are reached in Section 5.6.
5.2
Problem statement
We consider a general class of LTI discrete-time plants in the following form: Dz 1 y
t N
z 1 u
t
t
1
t
1
5:14
where u
t is the input, y
t is the output,
t is the exogenous disturbance,
t represents the unmodelled dynamics (to be speci®ed later), z 1 is the unit delay operator: N
z 1 n1 z
1
n2 z
D
z 1 1 d1 z
1
2
. . . nn z
. . . dn z
n
n
5:15
5:16
Remark 2.1 By using simple algebraic manipulations, measurement noise and input disturbance are easily incorporated into the model (5.14). In this case, y
t, u
t, and
t represent the measured output, computed input and (generalized) exogenous disturbance, respectively. For example, if a linear time-invariant discrete-time plant is described by y
z
N
z 1
u
z d
z q
z D
z 1
where d
z and q
z are the input disturbance and plant noise, respectively, the plant can be rewritten as D
z 1 y
z N
z 1 u
z
N
z 1 d
z D
z 1 q
z 1 Consequently, the exogenous input
z is N
z 1 d
z D
z 1 q
z 1 . We will denote by the vector of unknown parameters, i.e.
nn ; . . . ; n2 ; dn ; . . . ; d1 ; n1 T
5:17
Throughout the chapter, we will use the following nonminimal state-space description of the plant (5.14): x
t 1 A
x
t B
u
t E
t
t
5:18
Adaptive
87
where xt u
t
n 1 u
t
and the matrices A
, 0 . .. 0 0 A
0 .. . 0 nn
n 1 y
tT
1 j y
t
5:19
B
and E are constructed in a standard way 1 0 0 0 0 .. .. .. . . . 0 1 0 0 0 0 0 0 0 0
5:20 0 0 0 1 0 .. .. .. . . . 0 0 0 0 1 nn 1 n2 dn d2 d1
0 .. . 0
1 B
C ; 0 . .. 0 n1
0 .. . 0
0 E C 0 . .. 0
5:21
1
We also de®ne the regressor vector
t
x
t u
t
5:22
Then, (5.14) can be rewritten as y
t T
t
1
t
1
t
1
5:23
The following assumptions are used throughout this section: (A1) The order n of the nominal plant (excluding the unmodelled dynamics) is known. (A2) A compact set 2 2n , is known such that 2 :
88
(A3) The plant (5.14) without unmodelled dynamics (i.e. t 0) is stabilizable over . That is, for any 2 , there exists a linear time-invariant controller C
z 1 such that the closed loop system is exponentially stable. (A4) The exogenous disturbance is uniformly bounded, i.e. for all t0 2 N sup j
tj
5:24
tt0
for some known constant : (A5) The unmodelled dynamics is arbitrary subject to j
tj
t " sup t k jjx
kjj
5:25
0kt
for some constants " > 0 and 0 < 1 which represent the `size' and `decay rate' of the unmodelled dynamics, respectively. Remark 2.2 Assumption (A1) can be relaxed so that only an upper bound nmax is known. Assumption (A4) will be used in Sections 5.3±5.4 and will be relaxed to allow to be unknown in Sections 5.3.2 and 5.4.1 where an estimation scheme is given for . Remark 2.3 We note that the assumptions outlined above are quite standard and have been used in adaptive control to derive stability results for systems with unmodelled dynamics (see, e.g., [7, 16, 21, 30] for more details). The switching controller to be designed will be of the following form: u
t Ki
t x
t
5:26
where Ki
t is the control gain applied at time t, and i
t is the switching index at time t, taking value in a ®nite index set I. The objective of the control design is to determine the set of control gains KI fKi ; i 2 Ig
5:27
and an on-line switching algorithm for i
t so that the closed loop system will be `stable' in some sense. We note that switching controllers can be classi®ed according to the logic governing the process of switching. Here are some typical examples.
The switching index is de®ned as i
t 1 i
t i
t 1 1
if Gt 0 otherwise
5:28
where Gt is some appropriately chosen performance index. This type of
89
switching control is ®nitely convergent and based on an exhaustive search over the parameter space (see, e.g., [8], [9]).
The switching index is de®ned as it 1 it arg mini2I jei
tj
if t
s
t < td
otherwise
5:29
where s
t is the time of the most recent switching, td is a positive dwell time, and ei
t, 8i 2 I is a weighted prediction error computed for the ith nominal system. This type of switching control has been extensively studied recently by a number of researchers (see, e.g., [25, 26]). The proof of the closed loop stability in this case is not dependent on ®nite convergence of the switching process, furthermore, supervisory switching control is not ®nitely convergent in general.
5.3
Direct localization principle
The switching algorithms to be used in this section are based on a localization technique. This technique, originally used in [39] for LTI plants, allows us to falsify incorrect controllers very rapidly while guaranteeing exponential stability of the closed loop system. In this section, we describe a direct localization principle (see, e.g., [40]) for LTI plants which is slightly dierent from [39] but is readily extended to LTV plants. The main idea behind this principle consists of simultaneous falsi®cation of potentially stabilizing controllers based explicitly on the model of the controlled plant. That implies the use of some eective mechanism of discarding controllers inconsistent with the measurements. The speci®c notion of stability to be used in this section is described below. De®nition 3.1 The system (5.14) satisfying (A1)±(A5) is said to be globally exponentially stabilized by the controller (5.26) if there exist constants M1 > 0, 0 < < 1, and a function M2
: R ! R with M2
0 0 such that jjx
tjj M1
t
t0
jjx
t0 jj M2
5:30
holds for all t0 0, x
t0 , 0, and
and
satisfying (A4)±(A5), respectively. The de®nition above yields exponential stability of the closed loop system provided that 0 and exponential attraction of the states to an origin centred ball whose radius is related to the magnitude of the exogenous disturbance. First, we decompose the parameter set to obtain a ®nite cover f i gLi1 which satis®es the following conditions:
90
Adaptive
(C1)
i , i 6 f g; i 1; . . . ; L: (C2) Li1 i . (C3) For each i 1; . . . ; L, let i and ri > 0 denote the `centre' and `radius' of
i , i.e. i 2 i and jj i jj ri for all 2 i . Then, there exist Ki , i 1; . . . ; L, such that jmax
A
B
Ki j < 1;
8 k
i k ri ;
i 1; . . . ; L:
5:31
Conditions (C1)±(C2) basically say that the uncertainty set is presented as a ®nite union of non-empty subsets while condition (C3) de®nes each subset i as being stabilizable by a single LTI controller Ki . It is well known that such a ®nite cover can be found under assumptions (A1)±(A3) (see, e.g., [8, 24, 25] for technical details and examples). More speci®cally, there exist (suciently large) L, (suciently small) ri , and suitable Ki , i 1; . . . ; L, such that (C1)±(C3) hold. Leaving apart the computational aspects of decomposing the uncertainty set satisfying conditions (C1)±(C3) we just note that decomposition can be conducted o-line, moreover, some additional technical assumptions (see, e.g., (C3 0 ) below) make the process of decomposing pretty trivial. The computational complexity of decomposing the uncertainty set, in general, depends on many factors including the `size' of the set, its dimension and `stabilizability' properties, and has to be evaluated on a case-by-case basis. The key observation used in the localization technique is the following fact: given any parameter vector 2 j and a control gain Ki
t for some i
t; j 1; . . . ; L. If i
t j, then it follows from that
y
t T
t jTj
t
1
1
t
y
tj rj jj
t
1
t
5:32
1
1jj
t
1
5:33
This observation leads to a simple localization scheme by elimination: If the above inequality is violated at any time instant, we know that the switching index i
t is wrong (i.e. i
t 6 j), so it can be eliminated. In identi®cation theory this concept is sometimes referred to as falsi®cation; see, e.g., a survey [15] and references therein. The unique feature of the localization technique comes from the fact that violation of (5.33) allows us not only to eliminate i
t from the set of possible controller indices, but many others. This is the key point! As a result, a correct controller can be found very quickly. We now describe the localization algorithm. Let I
t denote the set of `admissible' control gain indices at time t and initialize it to be I
t0 f1; 2; . . . ; Lg
5:34
Choose any initial switching index i
t0 2 I
t0 . For t > t0 , de®ne ^ f j :
5:33 holds; I
t
j 1; . . . ; Lg
Then, the localization algorithm is simply given by
5:35
Adaptive
It It
^ 1 \ I
t; 8t > t0
91
5:36
1
The switching index is updated by taking i
t 1 if t > t0 and i
t i
t any member of I
t otherwise
1 2 I
t
5:37
A simple geometrical interpretation of the localization algorithm (5.36) is given in Figure 5.2. One possible way to view the localization technique is to interpret it as family set identi®cation of a special type, that is, family set identi®cation conducted on a ®nite set of elements. Interpreted in this way the localization technique represents a signi®cant departure from traditional family set identi®cation ideas. Either strip depicted in Figure 5.2 contains only those elements which are consistent with the measurement of the input/output pair fy
t; u
t 1g. The high falsifying capability of the proposed algorithm observed in simulations can informally be explained in the following way. Let the index i
t be falsi®ed, then the discrete set of elements fi : i 2 I
tg consistent with all the past measurements is separated from the point i
t by one of the hyperplanes Tj
t
1 y
t rj jj
t
1jj
t
1
5:38
Tj
t
1 y
t
1jj
1
5:39
or rj jj
t
t
dividing the parameter space into two half-spaces. It is also clear that every element belonging to the half-space containing the point i
t is falsi®ed by the algorithm of localization (5.36) at the switching instant t. We note that the rigorous analysis of the problem of optimal localization conducted in Section 5.3.1 allows us to derive a guaranteed lower bound on the number of controllers falsi®ed at an arbitrary switching instant. A dierent non-identi®cation based interpretation of localization can be given in terms of the prediction errors ej jTj
t 1 y
tj, j 1; 2; . . . L computed for the entire set of `nominal' models. Thus, any model giving a large prediction error is falsi®ed. The following technical lemma describes the main properties of the algorithm of localization (5.36). Lemma 3.1 Given the uncertain system (5.14) satisfying assumptions (A1)± (A5), suppose the ®nite cover f i gLi1 of satis®es conditions (C1)±(C3). Then, the localization algorithm given in (5.34)±(5.37) applied to an LTI plant (5.14) possesses the following properties: (i) I
t 6 f g;
8t t0 :
1 In fact, we will see in Section 5.3.1 that there may be `clever' ways of selecting i
t when i
t 1 is falsi®ed.
92
Adaptive PARAMETER
SPACE
Θ i(t)
i
t
UNCERTAINTY SET
Figure 5.2
(ii) There exists a switching index j 2 I
t for all t t0 such that the closed loop system with u
t Kj x
t is globally exponentially stable. Proof The proof is trivial: suppose the parameter vector for the true plant is in j for some j 2 f1; . . . ; Lg. Then, the localization algorithm guarantees that ^ for all t. Hence, both (i) and (ii) hold. j 2 I
t To guarantee exponential stability of the closed loop system, we need a further property of the ®nite cover of . To explain this, we ®rst introduce the notion of quadratic stability [3]. De®nition 3.2 A given set of matrices fA
: 2 g is called quadratically stable if there exist symmetric positive-de®nite matrices H; Q such that AT
HA
H
Q;
82
5:40
It is obvious that the ®nite cover f i gLi1 of can always be made such that
Adaptive
93
each i is `small' enough for the corresponding family of the `closed-loop' matrices fA
B
Ki : 2 i g to be quadratically stable with some Ki . In view of the observation above, we replace the condition (C3) with the following: (C3 0 ) For each i 1; . . . ; L, let i and ri > 0 denote the `centre' and `radius' of
i , i.e. i 2 i and jj i jj ri for all 2 i . Then, there exist control gain matrices Ki , symmetric positive-de®nite matrices Hi and Qi , i 1; . . . ; L, and a positive number q such that
A
B
Ki T Hi
A
B
Ki 8 k
i k
ri q;
Hi
Qi ;
i 1; . . . ; L
5:41
Remark 3.1 Condition (C3 0 ) requires that every subset i obtained as a result of decomposition be quadratically stabilized by a single LTI controller. We also note that a ®nite cover which satis®es (C1)±(C2) and (C3 0 ) is guaranteed to exist. Moreover, Condition (C3 0 ) translated as one requiring the existence of a common quadratic Lyapunov function for any subset i further facilitates the process of decomposition. The following theorem contains the main result for the LTI case: Theorem 3.1 Given an LTI plant (5.14) satisfying assumptions (A1)±(A5). Let f i gLi1 be a ®nite cover of satisfying conditions (C1)±(C2) and (C3 0 ). Then, the localization algorithm given in (5.34)±(5.37) will guarantee the following properties when " (i.e. the `size' of unmodelled dynamics) is suciently small: (i) The closed loop system is globally -exponentially stable, i.e., there exist constants M1 > 0, 0 < < 1, and a function M2
: R ! R ; M2
0 0 such that jjx
tjj M1
t
t0
jjx
t0 jj M2
5:42
holds for all t t0 and x
t0 : (ii) The switching sequence fi
t0 ; i
t0 1; . . .g is ®nitely convergent, i.e. i
t const, 8 t t for some t . Proof
See Appendix A.
The proof of the theorem presented in Appendix A is based on the observation that between any two consecutive switchings the closed loop system behaves as an exponentially stable LTI system subject to small parametric perturbations and bounded exogenous disturbance. This is the key point oering a clear understanding of the control mechanisms. It follows from the proof of Theorem 3.1 that the constant M1 in the bound (5.42) is proportional to the total number of switchings made by the controller
94
while the parameter is dependent on the `stabilizability' property of the uncertainty set . This further emphasizes the importance of fast switching capabilities of the controller for achieving good transient performance.
!" # The localization scheme described above allows an arbitrary new switching index in It to be used when a switching occurs. That is, when the previous switching index it 1 is eliminated from the current index set I
t, any member of I
t can be used for i
t. The problem of optimal localization addresses the issue of optimal selection of the new switching index at each switching instant so that the set of admissible switching indices I
t is guaranteed to be pruned down as rapidly as possible. The problem of optimal localization is solved in this section in terms of the indices of localization de®ned below. In the following for notational convenience we drop when possible the index t from the description of the set of indices I
t. Also we make the technical assumption (A7) ri rj r; 8i; j 1; 2; . . . ; L; and some r > 0. For any set I f1; 2; . . . ; Lg, fi : i 2 Ig, a ®xed j 2 I and any z 6 0, z 2 2n , de®ne the function
z; j; jfi :
i
j T z 0;
i 2 Igj
5:43
where j j denotes the cardinal number of a set. Then ind
j ; min
z; j;
5:44
jjzjj1
will be referred to as the index of localization of the element j with respect to the set . Lemma 3.2 The index of localization ind
j ; represents a guaranteed lower bound on the number of indices discarded from the localization set I
t at the next switching instant provided that u
t Kj x
t. Proof Without loss of generality we assume that
t 1 is the next switching ^ 1, = I
t instant, and controller Kj is discarded. From (5.35) we have j 2 equivalently Tj
t > y
t 1
rj qjj
tjj
t 1 or Taking z
Tj
t < y
t 1
rj qjj
tjj
t 1
5:45
5:46
t=jj
tjj for (5.45), or z
t=jj
tjj for (5.46) and using
Adaptive
95
(5.43) we see that there are z; j; number of controller indices which do not ^ 1. We note that
t 6 0, because otherwise it is easy to see belong to It from (5.23) that there exists no element j 2 satisfying (5.45) or (5.46), and, consequently, switching is not possible. Since ind
j ;
z; j; , we conclude that there are at least ind
j ; number of controllers to be discarded at the switching instant
t 1. In terms of (5.44) the index of localization of the discrete set is de®ned as ind maxfind
j ; : j 2 Ig j
5:47
That is, ind , is the largest attainable lower bound on the number of controllers eliminated at the time of switching, assuming that the regressor vector can take any value. The structure of an optimal switching controller is described by u
t Ki
t x
t i
t 1 i
t iopt
t arg maxj find
j ;
t : j 2 I
tg
5:48 if i
t 1 2 I
t; otherwise
5:49
The problem of optimal localization reduces to determining the optimal control law, that is, specifying the switching index iopt
t at each time instant when switching has to be made. To solve this problem we introduce the notion of separable sets. De®nition 3.3 Given a ®nite set Rn and a subset J ; J is called a separable set of order k if (i) jJj k. (ii) co fJg co f set.
Jg fg where co fg stands for the convex hull of a
The main properties of separable sets are listed below: (a) A vertex of co fg is a separable set of order 1. (b) The order of a separable set k jj. (c) For each separable set J of order k; k > 1, there exists a set J 0 J such that J 0 is a separable set of order
k 1. Proof (a), (b) are obvious. To prove (c), we note that for each separable set J, there exists a hyperplane P separating J and J. Let ~ n be the normal direction of P. Move P along ~ n towards J until it hits J. Two cases are possible. Case 1: One vertex is in contact. In this case move P a bit further to pass the vertex. The remaining points in J form J 0 .
96
Adaptive
Case 2: Multiple vertices are in contact. One can always change n slightly so that still separates J and J, but there is only one vertex in contact with , and we are back to Case 1. k Lemma 3.3 Let K be the set of all separable sets of order k and k H Jk 2Kk Jk . Then,
ind 1 arg maxfk : Hk 6 g
5:50
k
Proof Follows immediately from De®nition 3.3 and the property of separable sets (c). Indeed, suppose that the index of localization satis®es the relation ind m > 1 arg maxfk : Hk 6 g
5:51
k
then there must exist an element j 2 , such that ind
j ; m, moreover = Hm 1 ; j 2
j 2
Hm
1
5:52
since otherwise, by de®nition of separable sets ind
j ; m 1. But it follows from (5.51) that Hm 1 fg. On the other hand by De®nition 3.3 and the properties of separable sets (b), (c) the index of localization of the set cannot be smaller than that given by (5.50). This concludes the proof. Denote by V
the set of vertices of co
. The complete solution to the problem is given by the following iterative algorithm. Algorithm A Step 1
Initialize k 1. Compute K1 ffg : 2 V
g:
Step 2
Set k k 1. Compute
Kk fJk Step 3
1
[ i : Jk
1
Kk 1 ;
i 2 V
Jk 1 ;
Jk
1
[ i is separableg
If Hk , then ind k, and stop, otherwise go to Step 2.
The properties of localization based switching control are summarized in the following theorem. Let subfg denote the set of subscripts of all the elements in fg. Theorem 3.2 (i) The solution to the problem of optimal localization may not be unique and is given by the set
5:53 Iopt subf Hm 1 g where
m ind 1 arg maxfk : Hk 6 g k
5:54
(ii) For any 0; " 0, the total number of switchings l made by the
Adaptive
97
optimal switching controller (5.48), (5.49) applied to an LTI plant (5.14) satis®es the relation l 1
ind
tp 2 4 L 1
5:55 p0
where tp ; p 0; 1; . . . ; l Proof that
1 denote the switching instants.
The proof of (i) follows directly from Lemma 3.3. To prove (ii) we note j
t1 j L
ind
t0 ;
j
t2 j j
t1 j
ind
t1 L
ind
t0
ind
t1
then j
tl j l L
l 1
ind
ti
i0
Since l 1 the result follows.
Algorithm A applied to an arbitrary localization set indicates that except for a very special case, namely, fj gj2I V
, localization with any choice of = V
will always result in elimination the switching index i
t such that i
t 2 of more than one controller at any switching instant. This is a remarkable feature distinguishing localization based switching controllers from conventional switching controllers. Moreover, a simple geometrical analysis (see, e.g., Figure 5.2) indicates that for `nicely' shaped uncertainty sets (for example, a convex ) and large L the index of localization is typically large, that is, ind
>> 1. Theorem 3.2 gives a complete theoretical solution to the problem of optimal localization formulated above in terms of indices of localization. However, it must be pointed out that the search for optimality in general is involved and may be computationally demanding. To alleviate potential computational diculties we propose one possible way of constructing a suboptimal switching controller. Algorithm B Step 1
Initialize k 1. Compute A1 V
.
Step 2
Set k k 1. Compute Ak Ak
Step 3
1
[ V
Ak 1
If Ak , then ind k, and stop, otherwise go to Step 2.
Algorithm B allows for a simple geometrical interpretation, namely, at each step a new set Ak is obtained recursively by adding the set of vertices of
98
Adaptive
Ak 1 . The simplicity of the proposed algorithm is explained by the fact that we no longer need to check the property of separability (see Step 2 in Algorithm A). The main property of the Algorithm B is presented in the following proposition. Proposition 3.3
The index of localization ind satis®es the inequality ind 1 arg max fk : Ak 6 g k
5:56
Proof The proof is very simple and follows from the fact that for any 2 , such that 2 = V
it is true that ind
; 2. By applying this rule recursively we obtain (5.56). Example 3.1 To illustrate the idea of optimal (suboptimal) localization we consider a simple localization set fj g5j1 depicted in Figure 5.3. We note that the point 5 is located exactly in the centre of the square
1 ; 2 ; 4 ; 3 . Applying Algorithm A to the set we obtain K1 ff1 g; f2 g; f3 g; f4 gg; K2 ff1 ; 2 g; f1 ; 3 g; f2 ; 4 g; f3 ; 4 gg; K3 ff1 ; 2 ; 5 g; f1 ; 3 ; 5 g; f3 ; 4 ; 5 g; f2 ; 4 ; 5 gg Since [J2K3 J we conclude that ind 3 and the optimal switching index is given by i
t 5. To compute a guaranteed lower bound on the index of 1
2 *
* 5 * 3
4 *
Figure 5.3
*
99
localizaton ind Algorithm B is used. We have A1 f1 ; 2 ; 3 ; 4 g; A2 f1 ; 2 ; 3 ; 4 ; 5 g therefore, ind 2. We note that in this particular example the optimal solution, that is, i
t 5, coincides with the suboptimal one. Remark 3.2 To deal with the problem of optimal (suboptimal) localization dierent simple heuristic procedures can be envisioned. For example, the following `geometric mean' algorithm of computing a new switching index is likely to perform well in practice, though it is quite dicult in general to obtain any guaranteed lower bounds on the indices of localization. At any switching instant t we choose
i =
tjk
5:57 i
t arg min jjj j
i2I
t
!"
$ In this section we further relax assumption (A4) to allow the disturbance bound to be unknown. That is, we replace (A4) with (A4 0 ) The exogenous disturbance is uniformly bounded sup j
tj
5:58
tt0
for some unknown constant . We further relax assumptions (A1)±(A5) by allowing parameters to be slowly varying. To this end we introduce the following additional assumption (A7) The uncertain parameters are allowed to have slow drifting described by jj
t
t
1jj ;
8 t > t0
5:59
for some constant > 0. Following the results presented in previous sections, we introduce a generalized localization algorithm to tackle the new diculty. The key feature of the This estimate starts with a small algorithm is the use of an on-line estimate of . (or zero) initial value, and is gradually increased when it is invalidated by the observations of the output. With the trade-o between a larger number of switchings and a higher complexity, the new localization algorithm guarantees qualitatively similar properties for the closed loop system as for the case of known disturbance bound.
100
Adaptive
be the estimate for at time t. De®ne Let t ^ t It; f j : jTj
t
t
1
1
t
y
tj
ri qk
t
1k
5:60
j 1; . . . ; Lg
1;
^
t That is, I
t; is the index set of parameter subsets which cannot be falsi®ed by any exogenous disturbance suptt0 j
tj
t 1. as Denote the most recent switching instant by s
t. We de®ne s
t and
t follows: s
t0 t0
5:61
0 0
t
s
t
1 otherwise
t
t
1
t
if
1
otherwise
s
t
t
5:62 if
t
t
ks
t 1
^
k I
k; fg
t
ks
t
^
k I
k;
and
t
1 f g
s
t td
and
t
5:63 s
t < td
5:64
where td is some positive integer representing a length of a moving time interval is conducted, is a small positive over which validation of a new estimate
t constant representing a steady state residual (to be clari®ed later), and
t is an integer function de®ned as follows: t ^
k
5:65 I
k; 1 6 f g; 2 N
t min : ks
t The main idea behind the estimation scheme presented above is as follows. At each time instant when the estimate
t 1 is invalidated, that is, t ^
k I
k; 1 fg we determine the least possible value 2 N which ks
t 1 guarantees that no exogenous disturbance suptt0 j
tj
t would have caused the falsi®cation of all the indices in the current localization set. This is done by recomputing the sequence of localization sets over the ®nite period of time s
t; t whose length is bounded from above by td . Since the is ®nite and for total number of switchings caused by the `wrong' estimate
t every suciently large interval of time the number of switchings due to slow parameter drifting can be made arbitrarily small by decreasing the rate of slow parameter drifting it is always possible to choose a suciently large td which would guarantee global stability of the system. The algorithm of localization is modi®ed as follows: I
t \tks
t
^
1 I
k;
k
But the switching index i
t is still de®ned as in (5.37).
5:66
Adaptive
101
The key properties of the algorithm above are given as follows: Theorem 3.4 For any constant > 0, there exist a parameter drifting bound > 0, a `size' of unmodelled dynamics " > 0 (both suciently small), and an integer td (suciently large), such that the localization algorithm described above, when applied to the plant (5.14) with assumptions (A1)±(A3), (A4 ) and (A7), possesses the following properties: (1) It 6 fg for all t t0 . . (2) suptt0
t Subsequently, the following properties hold: (3) The closed loop system is globally
-exponentially stable, i.e. there exist constants M1 > 0, 0 < < 1, and a function M2
: R ! R with M2
0 0 such that jjx
tjj M1
t
t0
jjx
t0 jj M2
5:67
holds for all t t0 and x
t0 . (4) The switching sequence fi
t0 ; i
t0 1; . . .g is ®nitely convergent, i.e. i
t const, 8 t t for some t if the uncertain parameters are constant. Proof
See Appendix B.
We note that even though the value can be arbitrarily chosen, the estimate of the disturbance bound,
t, can theoretically be larger than by the margin . Consequently, the state is only guaranteed to converge to a residual set slightly larger than what is given in Theorem 3.4. Our simulation results very likely converges to a value substantially smaller than . indicate that
t Nevertheless, there are cases where
t exceeds . One possible solution is to reduce the value of . However, a small may imply a large number of potential switchings.
5.4
Indirect localization principle
The idea of indirect localization was ®rst proposed in [39] and is based on the use of a specially constructed performance criterion as opposed to direct localization considered in the previous section. To this end the output of the plant is replaced by some auxiliary output observation which is subsequently used for the purpose of model falsi®cation. The notion of `stabilizing sets' introduced below is central in the proposed indirect localization scheme. We ®rst de®ne an auxiliary output, z
t, as z
t Cx
t;
C T 2 R2n
1
5:68
102
Adaptive
and the inclusion:
t : jz
tj kx
t
1k c0
5:69
De®nition 4.1 I t is said to be a stabilizing inclusion of the system (5.18) if I t being satis®ed for all t > t0 and boundedness of
t;
t 2 l1 , implies boundedness of the state, x
t, and in particular, there exist 0 ; 0 and 2
0; 1 such that: kx
tk 0 t
t0
kx
t0 k 0 kt k`1
Remark 4.1 Note that the inclusion, I t is transformed into a discrete-time sliding hyperplane [11] as ! 0; c0 ! 0. In contrast with conventional discrete-time sliding mode control we explicitly de®ne an admissible vicinity around the sliding hyperplane by specifying the values 0 and c0 0. De®nition 4.2 stabilizable if
The uncertain system (5.18) is said to be globally
C; -
(1) I t is a stabilizing inclusion of the system (5.18), and (2) there exists a control, u
t Kx
t, such that after a ®nite time, I t is satis®ed. We will show below that stabilizing sets can be eectively used in the process of localization. Before we proceed further we need some preliminary results. Assume for simplicity that
t 0. The case
t 6 0 is analysed similarly, provided " is suciently small. Lemma 4.1 Let suptt0 j
tj < 1; CB > 0. Then there exists a c0 such that the system (5.18) is globally
C; 0-stabilizable if and only if where
Proof
jmax
PAj < 1 P I
CB 1 BC
5:70
5:71
First, suppose that (5.70) is violated, that is jmax
PAj 1
5:72
We now show that I t is not a stabilizing inclusion for any c0 > 0. To do this, we take
t 0, and u
t
CB 1 CAx
t. With this control we note from (5.18) that z
t 0 for t > 0, and so for any c0 > 0; z
t satis®es (5.69). The equation for the closed loop system takes the form x
t 1 Ax
t Bu
t PAx
t
5:73
which is not exponentially stable. Therefore, (5.72) implies that there is no c0 such that I t is a stabilizing inclusion. We now establish the converse. Suppose
Adaptive
103
(5.70) is satis®ed. Then we can rewrite (5.18) as: xt 1 PAx
t Bu
t PAx
t
1
BCAx
t E
t CB
B
z
t 1 CB
5:74
CE
t E
t
From (5.74) it is clear that if z and are bounded, then in view of (5.70), x
t is bounded. Therefore, I t is a stabilizing inclusion for any c0 . Finally, we take the control 1
CAx
t
5:75 u
t CB which gives z
t 1 CE
t
5:76 Therefore, for c0 jCEj supt j
tj; I t is satis®ed for all t > 0, and the proof is complete. Remark 4.2 The control, (5.75), is a `one step ahead' control on the auxiliary output, z
t. It then follows that the stability condition (5.70), (5.71) is equivalent to the condition that C
zI A 1 B be relative degree one, and minimum phase. Remark 4.3 If the original plant transfer function, (5.14), is known to be minimum phase, and relative degree one then it suces to take C E T , and the system is then c0 stabilizable for any c0 0. If the original plant transfer function is non-minimum phase, then let: C f 0 ; f 1 . . . f n 2 ; g0 ; g 1 . . . gn 1
5:77
The transfer function from u
t, via (5.14) to z
t is then: z
t F
qu
t G
qy
t
5:78 D
qF
q G
qN
q u
t D
q and G
q
g0 g1 q . . . where F
q f0 f1 q fn 2 qn 2 gn 1 qn 1 . Therefore, for a non minimum phase plant, knowledge of a C such that I t is a stabilizing inclusion is equivalent to knowledge of a (possible improper) controller fu
t G
q=F
qy
tg which stabilizes the system. Because we are dealing with discrete-time systems, it is not clear whether this corresponds to knowledge of a proper, stabilizing controller for the set. Remark 4.4 Because of the robustness properties of exponentially stable linear time invariant systems, Lemma 4.1 can easily be generalized to include nonzero, but suciently small .
104
Adaptive
Lemma 4.2 Any which satis®es assumptions (A2) and (A3) has a ®nite decomposition into compact sets:
L
5:79
1
such that for each l, there exists a Cl ; l and c0;l such that, for all A; B 2 l ; I t is a stabilizing inclusion, and Cl B has constant sign. Proof (Outline) It is well known that (see, e.g., [8]) that has a ®nite decomposition into sets stabilized by a ®xed controller. From Remark 4.3, the requirements for knowledge of a Cl such that I
Cl ; ; is a stabilizing set on are less stringent than knowledge of a stabilizing controller for the set l . We now introduce our control method, including the method of localization for determining which controller to use. The ®rst case we consider is the simplest case where there is a single set to consider. Case 1: L 1 (sign of CB known) This case covers a class of minimum phase plants, plus also certain classes of nonminimum phase plants. For L 1 we have: s
1
i
5:80 i1
For i 1 . . . s we de®ne a control law: u
ti
Ki x
tX
1 CAi x
t CBi
5:81
where the plant model, Ai ; Bi is in the set i . We require knowledge of a such that: CB kC
A Ai
5:82 k ; 8i; 8
A; B 2 i CBi and I t is a stabilizing inclusion on i for all i. Note that for any bounded , for which we can ®nd a single C which gives C
zI A 1 B minimum phase and relative degree one we can always ®nd, for s large enough, a with the required properties (see, e.g., [8]). At any time t > 0, the auxiliary output z
t 1i which would have resulted if we applied u
ti Ki x
t to the true plant is, using (5.18). z
t 1i XCAx
t CBu
ti CE
t z
t 1 CB u
t u
ti
5:83
Adaptive
105
Note that if the true plant is in the set i , then from (5.83) and (5.81) CB i zt 1 C A Ai x
t CE
t
5:84 CBi and, therefore, if the true plant is in i , then from (5.82), and with c0 jCE j jz
t 1i j kx
tk c0
5:85
Our proposed control algorithm for Case 1 is as follows (where, without loss of generality, we take CB > 0). Algorithm C 1.1 Initialization De®ne
S0 f1; 2; . . . ; sg
5:86
1.2 If t > 0, If z
t > jx
t 1j c0 then set St St 1 fk; . . . ; js 1 ; js g If z
t > jx
t 1j c0 then set St St 1 fj1 ; j2 ; . . . ; kg otherwise, St St 1 . where k; j1 . . . js and s are integers from the previous time instant (see 1.4, 1.5). 1.3 If t > 0, For all i 2 St , compute u
ti as in (5.81). 1.4 Order u
ti ; i 2 St such that: u
t j1 u
t j2 . . . u
t js
5:87
1.5 Apply the `median' control: u
t u
tk where k jbs=2c , 1.6 Then wait for the next sample and return to 1.2.
5:88
We then have the following stability result for this control algorithm. Theorem 4.1 The control algorithm, (5.86)±(5.88), applied to a plant where C is known, and where the decomposition (5.80) has the properties that (5.82) is satis®ed and I t is a stabilizing inclusion, has the following properties: (a) The inclusion:
I t : jz
tj kx
t
1k c0
5:89
is violated no more than N blog2
sc times, and (b) All signals in the closed loop system are bounded. In particular, there exist constants ; < 1; 2
0; 1 such that all trajectories satisfy, for any t0 ; T > 0
5:90 kx
t0 Tk T kx
t0 k
106
Adaptive
Proof (a) Suppose at time t 1, (5.89) is violated. This can occur in one of two ways which we consider separately: (i) z
t 1 z
t 1it > kx
tk c0
5:91 In this case, because of the ordering of u
ti in (5.87), and the de®nition of z
t 1i in (5.83), then z
t 1i > kx
tk c0 for all i 2 fk; . . . ; js 1 ; js g (ii) z
t 1 z
t 1it < In this case
z
t 1i
0 such that (5.89) is violated not more than once in the interval,
t0 ; t0 T, then
kx
t0 Tk 1 T kx
t0 k 1 Also, we can show that with 2
5:98
0 1 3 1 20 , and 2 0 0
Adaptive
107
provided (5.89) is not violated more than
0 2 0
1 0 , 1 0 twice in the interval t0 ; t0 T, then kx
t0 Tk 2 T kx
t0 k 2
5:99
Repeating this style of argument leads to the conclusion that with N N 1
0 0 N 0 0 ; N
0 N
0
0 1 then if there are not more than N switches in t0 ; t0 T, then kx
t0 Tk N T kx
t0 k N
5:100
The desired result follows from (a) since we know that there are at most N log2
s times at which (5.89) is violated. Case 2: L > 1 Suppose that we do not know a single C such that I t is a stabilizing inclusion, and CB is of known sign, then using ®nite covering ideas [8], as in Remark 4.3 let L sl L l
lm
5:101
l1
l1 m1
where for each l, we know Cl ; l ; cl0 such that I t is a stabilizing inclusion on
l and the sign of
Cl B is constant for all plants in l . At this point, one might be tempted to apply localization, as previously de®ned, on the sets l individually and switch from l should the set of valid indices, S l , become empty. Unfortunately, this procedure cannot be guaranteed to work. In particular, if l does not contain the true plant, I t need not be a stabilizing inclusion, and so divergence of the states may occur without violating (5.89). To alleviate this problem, we use the exponential stability result, (5.90), in our subsequent development. Algorithm D We initialize t
i 0; R0 f1; 2; . . . ; Lg and take any l0 2 R0 . We then perform localization on l , with the following additional2 steps: If at any time kx
tk > t t
i kx
t
ik
5:102 (where ; ; are the appropriate constants for l from Theorem 4.1), then we set Sl fg. If at any time t; S l becomes empty, we set Rt Rt 1 flg; t
i t, and we take a new l from Rt . 2 In fact, we can localize simultaneously within other O i ; i 6 l: however, for simplicity and brevity we analyse only the case where we localize in one set at a time.
108
With these modi®cations, it is clear that Theorem 4.1 can be extended to cover this case as well: Corollary 4.1 The control algorithm (5.86)±(5.88) with the above modi®cations applied to a plant with decomposition as in (5.101) satis®es: (a) There are no more than: L 1 Ll1 blog2
sl c instants such that jz
t 1lt j lt kx
tk c0;lt
5:103
(where lt denotes the value of l at time t). (b) All signals in the closed loop are bounded. In particular, there exist < 1; 2
0; 1 such that for any t0 ; T > 0 constants ; T kx
t0 k kx
t0 Tk Proof
5:104
Follows from Theorem 4.1.
!%
$ In the previous section the problem of indirect localization based switching control for linear uncertain plants was considered assuming that the level of the generalized exogenous disturbance
t was known. This is equivalent to knowing some upper bound on
t. The ¯exibility of the proposed adaptive scheme allows for simple extension covering the case of exogenous disturbances of unknown magnitude. This can be done in the way similar to that considered in Section 5.3.2. Omitting the details we just make the following useful observation. The control law described by Algorithms C and D is well de®ned, that is, Rt 6 fg for all t t0 if cl0 suptt0 jCl E
tj; 8l 1; . . . ; L. This is the key point allowing us to construct an algorithm of on-line identi®cation of the parameters cl0 ; l 1; . . . ; L.
5.5
Simulation examples
Extensive simulations conducted for a wide range of LTI, LTV and nonlinear systems demonstrate the rapid falsi®cation capabilities of the proposed method. We summarize some interesting features of the localization technique observed in simulations which are of great practical importance. (i) Falsi®cation capabilities of the algorithm of localization do not appear to be sensitive to the switching index update rule. One potential implication of this observation is as follows. If not otherwise speci®ed any choice of a new switching index is admissible and will most likely lead to good transient performance; (ii) The speed of localization does not appear to be closely related to the total
Adaptive
5
400
output
switching index
600
200 0 0
control
5
109
50 (a)
100
50 (c)
100
0
−5 0
50 (b)
100
0
−5 0
Figure 5.4 &
number of ®xed controllers obtained as a result of decomposition. The practical implication of this observation (combined with the quadratic stability assumption) is that decomposition of the uncertainty set can be conducted in a straightforward way employing, for example, a uniform lattice which produces subsets i , i 1; 2; . . . ; L of an equal size. Example 5.1 Consider the following family of unstable (possibly nonminimum phase) LTV plants: yt 1:2yt
1
1:22y
t
2 b1
tu
t
1 b2
tu
t
2
t
5:105
where the exogenous disturbance
t is uniformly distributed on the interval 0:1; 0:1, and b1
t and b2
t are uncertain parameters. We deal with two cases which correspond to constant parameters and large-size jumps in the values of the parameters. Case 1: Constant parameters The a priori uncertainty bounds are given by b1
t 2 1:6;
0:15 [ 0:15; 1:6;
b2
t 2 2;
1 [ 1; 2
5:106
i.e. f 1:6; 0:15 [ 0:15; 1:6 2; 1 [ 1; 2g. To meet the requirements of the localization technique, we decompose into 600 nonintersecting subsets with their centres i
b1i ; b2i ; i 1; . . . ; 600 corresponding to
Adaptive 20
400
0
output
600
200 0 0
control
50
50 (a)
100
parameter b2
2
−40 0 2
0
−50 0
−20
parameter b1
switching index
110
50 (c)
100
50 (e)
100
50 (b)
100
50 (d)
100
0
−2 0
0
−2 0
Figure 5.5 & ' (
b1i 2 f 1:6; b2i 2 f 2;
1:5; . . . ; 0:3; 1:9; . . . ; 1:1;
0:2; 0:2; 0:3; . . . ; 1:5; 1:6g 1; 1; 1:1; . . . ; 1:9; 2g
respectively. Figures 5.4(a)±(c) illustrate the case where is constant. The switching sequence fi
1; i
2; . . .g depicted in Figure 5.4(a) indicates a remarkable speed of localization. Case 2: Parameter jumps The results of localization on the ®nite set fi g600 i1 are presented in Figures 5.5(a)±(e). Random abrupt changes in the values of the plant parameters occur every 7 steps. In both cases above the algorithm of localization in Section 5.2 is used. However, in the latter case the algorithm of localization is appropriately modi®ed. Namely, I
t is updated as follows ^ ^ I
t 1 \ I
t if I
t 1 \ I
t 6 fg
5:107 I
t ^ I
t otherwise
Adaptive
111
Once (or if) the switching controller, based on (5.107) has falsi®ed every index in the localization set it disregards all the previous measurements, and the process of localization continues (see [40] for details). In the example above a pole placement technique was used to compute the set of the controller gains fKi g600 i1 . The poles of the nominal closed loop system were chosen to be (0, 0.07, 0.1). Example 5.2 Here we present an example of indirect localization considered in Section 5.4. The model of a third order unstable discrete-time system is given by
5:108 y
t 1 a1 y
t a2 y
t 1 a3 y
t 2 u
t
t where a1 ; a2 ; a3 are unknown constant parameters, and
t 0 sin
0:9t represents exogenous disturbance. The a priori uncertainty bounds are given by a1 2 1:6; 0:1 [ 0:1; 1:6;
b2 2 1:6; 0:1 [ 0:1; 1:6;
a3 2 0:1; 1:6
5:109
i.e. f 1:6; 0:1 [ 0:1; 1:6 1:6; 0:1[0:1; 1:60:1; 1:6g. Choosing 4
Output
2 0 −2 −4 0
20
40
60
80
100 (a)
120
140
160
180
200
20
40
60
80
100 (b)
120
140
160
180
200
6
Control
4 2 0 −2 −4 −6 0
Figure 5.6
112
Adaptive
the vector C and the stabilizing set as prescribed in Section 5.4, we obtain : jz
t 1j jjx
tjj c0
5:110
where C
0; 0; 1 and 0:6. We decompose into 256 nonintersecting subsets with their centres i
a1i ; a2i ; a3i ; i 1; . . . ; 256 corresponding to a1i 2 f 0:3; 0:7; 1:1; 1:5; 0:3; 0:7; 1:1; 1:5g
5:111
a2i 2 f 0:3; 0:7; 1:1; 1:5; 0:3; 0:7; 1:1; 1:5g
5:112
a3i 2 f0:3; 0:7; 1:1; 1:5g
5:113
respectively. This allows us to compute the set of controller gains fKi g256 i1 , Ki
k1i ; k2i ; k3i . Each element of the gain vector kij , i 2 f1; 2; 3g, j 2 f1; . . . ; 256g takes values in the sets (5.111), (5.112), (5.113), respectively. The results of simulation with 0 0:1, a1 1:1, a2 0:7, a3 1:4, are presented in Figure 5.6(a)±(b). Algorithm C has been used for this study.
5.6
Conclusions
In this chapter we have presented a new uni®ed switching control based approach to adaptive stabilization of parametrically uncertain discrete-time systems. Our approach is based on a localization method which is conceptually dierent from the existing switching adaptive schemes and relies on on-line simultaneous falsi®cation of incorrect controllers. It allows slow parameter drifting, infrequent large parameter jumps and unknown bound on exogenous disturbance. The unique feature of localization based switching adaptive control distinguishing it from conventional adaptive switching controllers is its rapid model falsi®cation capabilities. In the LTI case this is manifested in the ability of the switching controller to quickly converge to a suitable stabilizing controller. We believe that the approach presented in this chapter is the ®rst design of a falsi®cation based switching controller which is applicable to a wide class of linear time-invariant and time-varying systems and which exhibits good transient performance.
Appendix A Proof of Theorem 3.1 First we note that it follows from Lemma 3.1 and the switching index update rule (5.37) that the total number of switchings made by the controller is ®nite. Let ft1 ; t2 ; . . . ; tl g be a ®nite set of switching instants. By virtue of (5.31)±(5.33) the behaviour of the closed loop system between any two
Adaptive
113
consecutive switching instants ts ; tj ; 1 s; j l; tj ts is described by x
t 1
A
B
Ki
ts x
t E
t
t
A
i
ts B
i
ts Ki
ts x
t E
t
5:114
where j
tj ri
ts jj
tjj
t. Therefore, taking into account the structure of the parameter dependent matrices A
and B
, namely the fact that only the last rows of A
and B
depend on the last equation can be rewritten as ^ x
t 1
A
i
ts
t B
i
ts
tKi
ts x
t E
t
5:115
^ for some
t : jj
tjj ri
ts q and j
tj
t. This is a direct consequence of the fact that the last equation in (5.114) can be rewritten as y
t 1 Ti
ts
t
t and that maxjjjj1 jjT
tjj jj
tjj holds for any
t. By De®nition 3.2 and condition (C3 0 ) the system (5.115) is quad^ 0 and ts being ®xed; moreover, there exists a positive ratically stable with
t T de®nite matrix Hts Hts such that Pts
max
jj
tjjri
ts q
jjA
i
ts
t B
i
ts
tKi
ts jjHts < 1
5:116
Here jjxjjH
xT Hx1=2 and for any matrix A 2 nn , jjAjjH denotes the corresponding induced matrix norm. The equation (5.115) along with the property of quadratic stability guarantee that between any two consecutive switchings the closed loop system behaves as an exponentially stable LTI system subject to small parametric perturbations
t and bounded dis^ and this property holds regardless of the possible evolution of turbance
t the plant parameters. This is the key point making the rest of the proof transparent. Assume temporarily that
t 0, then it follows from (5.115), (5.116) that jjx
ts 1jjHts Pts jjx
ts jjHts ^ts
5:117
jjx
ts 2jjHts P2ts jjx
ts jjHts
Pts 1^ts
5:118
... jjx
ts kjjHts Pkts jjx
ts jjHts ^
k
Pits
1
5:119
i1
jjx
ts kjj
max
Hts =min
Hts 1=2 Pkts jjx
ts jj ^ts
k
Pits 1 =min
Hts 1=2
i1
where ^ts maxjj jjEjjHts .
5:120
114
Adaptive
Denote M max
max
Hts =min
Hts 1=2 ;
max Pi < 1;
t1 itl
max ^i =
min
Hi 1=2 M
t1 itl
1
Pij
1
t1 itl
0 for all x 6 0. It is said to be proper, or radially unbounded if V
x ! 1 as jxj ! 1. A function R ! R is said to be of class K if it is continuous, strictly increasing and
0 0. It is of class K1 if, in addition, it is proper. A function R R ! R is of class KL if, for each ®xed t, the function
; t is of class K and, for each ®xed s, the function
s; is decreasing and tends to zero at in®nity. 6.2.2 Passivity and feedback passivation We begin with a brief review of basic de®nitions and properties related to passive systems. Then we present several more recent results about feedback equivalence of a nonlinear system to a passive system via static state feedback. The interested reader is referred to [50, 12, 2] and references therein for the details. 6.2.2.1 Passivity and passive systems Consider a nonlinear control system with outputs: x f
x G
xu
6:1
y h
x
6:2
where x 2 Rn , u ; y 2 Rm , f Rn ! Rn , G Rn ! Rnm and h Rn ! Rm . Assume that these functions are locally Lipschitz with f
0 0 and h
0 0.
122
Adaptive nonlinear control: passivation and small gain techniques
De®nition 2.1 (Passivity) A system (6.1)±(6.2) is said to be C r -passive if there exists a C r storage function V Rn ! R , with V
0 0, such that, for all admissible inputs u, all initial conditions xo and all t 0 t yT
su
s ds
6:3 V
x
t V
xo 0
De®nition 2.2 (Strict passivity) A system (6.1)±(6.2) is said to be output strictly C r -passive if there exists a C r storage function V Rn ! R , with V
0 0, and a positive de®nite function S1
y such that, for all admissible inputs u, all initial conditions xo and all t 0 t t T o S1
y
s ds
6:4 y
su
s ds V
x
t V
x 0
0
If (6.4) holds with a positive de®nite function S2
x, i.e. t t V
x
t V
xo yT
su
s ds S2
x
s ds 0
6:5
0
then the system is said to be (state) strictly C r -passive. The passivity and strict passivity properties of a nonlinear system can be tested by the following ubiquitous nonlinear KYP lemma. Lemma 2.1 [10] A system (6.1)±(6.2) is (resp. strictly) C 1 -passive if and only if there exists a C 1 non-negative function V
x, with V
0 0, such that Lf V
x is negative semi-de®nite (resp. negative de®nite) and Lg V
x h
x for all x 2 Rn . A fundamental property of a passive system is the Lyapunov stability of its unforced system (i.e. u 0 in (6.1)±(6.2)) and its reduced system with zerooutput (i.e. y 0 in (6.1)). More interestingly, under a zero-state detectability condition, passive systems are asymptotically stabilizable by static output feedback. We ®rst recall the de®nition of the zero-state detectability. De®nition 2.3 (Zero-state detectability) A system of the form (6.1)±(6.2) is locally zero-state detectable if there exists a neighbourhood N of x 0 such that for any initial condition x
0 2 N , u0
y0
)
x
t 0
t!1
6:6
It is (globally) zero-state detectable if the property (6.6) holds with N Rn . Theorem 2.1 [11], [2] Assume that system (6.1)±(6.2) is C 0 -passive with a positive de®nite storage function V and that it is locally zero-state detectable. Then, for any `®rst-sector third-sector' function , that is, a continuous
Adaptive Control Systems
123
function such that yT
y > 0 if y 6 0 u
6:7
y
is an asymptotically stabilizing controller for the system (6.1)±(6.2). Furthermore, if V is proper and (6.1)±(6.2) is (globally) zero-state detectable, the equilibrium x 0 of the closed loop system (6.1)±(6.7) is globally asymptotically stable (GAS). 6.2.2.2 Feedback passivation As seen in the preceding subsection and demonstrated by numerous researchers, passive systems enjoy many desirable properties which turn out to be very useful for practical control systems design. Naturally, this motivated people to address the feedback passivation issue. That is, when can a nonlinear control system in the form (6.1)±(6.2) be rendered passive via a state feedback transformation? This question had remained open until [2] where Byrnes, Isidori and Willems nicely provided a rather complete answer by means of dierential geometric systems theory. De®nition 2.4 (Passivation) A system (6.1)±(6.2) is said to be feedback (strictly) Cr -passive if there exists a change of feedback law u 1
x 2
xv
6:8 r
such that the system (6.1)±(6.2)-(6.8) with new input v is (strictly) C -passive. Theorem 2.2 [2] normal form
Consider a nonlinear system (6.1)±(6.2) having a global z_ q
z; y
6:9
y_ b
z; y a
z; yu
where a
is globally invertible. Then system (6.1)±(6.2), or (6.9) is globally feedback equivalent to a (resp. strictly) C 2 -passive system with a positive de®nite storage function, if and only if it is globally weakly minimum phase (resp. globally minimum phase). As an application of this important result to the global stabilization of cascaded nonlinear systems, it was shown in [2] that several previous stabilization schemes via dierent approaches can be uni®ed. In particular, combining Theorems 2.1 and 2.2, the following general result can be established. Theorem 2.3 [2, 34] Consider a cascaded nonlinear system of the form _ f0
f1
; xy x_ f
x g
xu y h
x
6:10
124
Adaptive nonlinear control: passivation and small gain techniques
Suppose _ f0
is GAS at 0. It is also assumed that ff ; g; hg is zerostate detectable and Cr -passive with a positive de®nite and proper storage function (r 1. Then system (6.10) is GAS by smooth state feedback. It should be noted that corresponding local results also hold. 6.2.3 Nonlinear gain The concept of nonlinear gain has recently been brought to the literature from two apparently dierent routes: the state-space approach [42, 44, 45, 46] and the input±output approach [39, 9, 30]. The special case of linear or ane gain had previously been extensively used in the stability theory of interconnected systems [5, 11, 33]. In Section 6.4, this notion will play a key role in addressing the robust adaptive control problem for parametric-strict-feedback systems in the presence of nonlinear dynamic uncertainties. In what follows, we limit ourselves to state-space systems. We begin with the introduction of Sontag's input-to-state stability (ISS) concept in which the role of initial conditions are made explicit. Then, we give an output version of this notion to allow for a broader set of controlled dynamical systems. Recent applications of these notions to establish nonlinear small gain theorems for interconnected feedback systems are recalled in the subsection 6.2.3.2. 6.2.3.1 Input-to-state stable systems Consider the general time-varying dynamical system with outputs x_ f
t; x; u y h
t; x; u
6:11
where x 2 Rn is the state, u 2 Rm1 is the control input and y 2 Rm2 is the system output. Notice that many dynamical systems subject to exogenous disturbances can be described by a dierential equation of the form (6.11). Roughly speaking, the property of input-to-state stability (ISS) says that the ultimate bound of the system trajectories depends only on the magnitude of the control input u and that the zero-input ISS system is globally uniformly asymptotically stable (GUAS) at the origin. More precisely, De®nition 2.5 A system of the form (6.11) is said to be input-to-state stable (ISS) if for any initial conditions t0 0 and x
t0 and for any measurable essentially bounded control function u, the corresponding solution x
t exists for each t t0 and satis®es jx
tj
jx
t0 j; t
t0
kut0 ;t k
where is a class KL-function and is a class K-function.
6:12
Adaptive Control Systems
125
An output version of the ISS property was given in [21, 16] and is recalled as follows. De®nition 2.6 A system of the form (6.11) is said to be input-to-output practically stable (IOpS) if for any initial conditions t0 0 and x
t0 and for any measurable essentially bounded control function u, the corresponding solution x
t exists for each t t0 and satis®es jy
tj
jx
t0 j; t
t0
kut0 ;t k d
6:13
where is a class KL-function, is a class K-function and d is a non-negative constant. If d 0 in (6.13), the IOpS property becomes IOS (input-to-output stability). Moreover, the system (6.11) is input-to-state practically stable (ISpS) if (6.13) holds with y x. An obvious property of an IOpS system of form (6.11) is that the system is bounded-input bounded-output (BIBO) stable. In addition, for any IOS system (6.11), a converging input yields a converging output. Among various characterizations of the ISS property is the notion of an ISSLyapunov function which was introduced by means of some dierential dissipation inequality [46], [28]. In Section 6.4, we consider a class of uncertain systems with persistently exciting disturbances. As a consequence, an extension of the ISS-Lyapunov function notion turns out to be necessary. De®nition 2.7 A smooth function V is said to be an ISpS-Lyapunov function for system (6.11) if . V is proper and positive de®nite, that is, there exist functions 1 , 2 of class K1 such that 1
jxj V
x 2
jxj; 8 x 2 Rn
6:14 . there exist a positive-de®nite function , a class K-function and a nonnegative constant c such that the following implication holds:
fjxj
juj cg )
@V
x f
t; x; u @x
jxj
6:15
When (6.15) holds with c 0, V is called an ISS-Lyapunov function for system (6.11) as in [28]. With the help of arguments in [46, 47, 28], it is not hard to prove the following fact. Fact: If a system of form (6.11) has an ISpS- (resp. ISS-) Lyapunov function V, then the system is ISpS (resp. ISS). 6.2.3.2 Nonlinear small gain theorems In this subsection, we recall two nonlinear versions of the classical small
126
Adaptive nonlinear control: passivation and small gain techniques
(®nite-) gain theorem [5] which were recently proposed in the work [21, 20, 16]. They will be used to construct robust adaptive controllers and develop the stability analysis in Section 6.4. The interested reader is referred to [4, 6, 20, 21, 49] and references therein for applications of the nonlinear small gain theorems to several feedback control problems. Consider the general interconnected time-varying nonlinear system x_ 1 f1
t x1 y2 u1
y1 h1
t x1 y2 u1
6:16
x_ 2 f2
t; x2 ; y1 ; u2 ;
y2 h2
t; x2 ; y1 ; u2
6:17
where, for i 1; 2, xi 2 Rni , ui 2 Rm1i and yi 2 Rm2i , and fi , hi are C1 in all their arguments. It is assumed that there is a unique C1 solution to the algebraic loop introduced by the interconnection functions h1 , h2 . !"orem 2.4 [21, 16] Assume that the subsystems (6.16) and (6.17) are IOpS in the sense that, for all 0 t0 t, jy1
tj 1
jx1
t0 j; t t0 1y ky2t0 ;t k 1u ku1t0 ;t k d1
6:18 jy2
tj 2
jx2
t0 j; t t0 2y ky1t0 ;t k 2u ku2t0 ;t k d2
Assume further that the subsystems (6.16) and (6.17) are also ISpS (resp. ISS). If there exist two class K1 -functions 1 and 2 and a nonnegative real number sl satisfying: )
I 2 2y
I 1 1y
s s 8s sl
6:19
I 1 1y
I 2 2y
s s then the interconnected system (6.16)±(6.17) with
y1 ; y2 as output is ISpS (resp. ISS) and IOpS
resp. IOS when sl di 0 for i 1; 2. Motivated by the above Fact, we consider properties of the feedback system derived from existence of ISpS-Lyapunov functions on each subsystem. Following step by step the proof of Theorem 3.1 in [20], we can prove the following result. Theorem 2.5 Consider the interconnected system (6.16)±(6.17) with y1 x1 and y2 x2 . Assume that, for i 1; 2, the xi -subsystem has an ISpS-Lyapunov function Vi satisfying the properties (1) there exist class K1 -functions i1 and i2 such that i1
jxi j Vi
xi i2
jxi j;
8 xi 2 Rni
6:20
(2) there exist class K1 -functions i , class K-functions i ; i and some constant ci 0 so that V1
x1 f1
V2
x2 ; 1
ju1 j c1 g implies @V1
x1 f1
t; x1 ; x2 ; u1 @x1
1
V1
x1
6:21
Adaptive Control Systems
127
and V2
x2 f2
V1
x1 ; 2
ju2 j c2 g implies @V2
x2 f2
t; x1 ; x2 ; u2 @x2
2
V2
x2
6:22
If there exists some c0 0 such that 1 2
r < r;
8 r > c0 ;
6:23
then the interconnected system (6.16)±(6.17) is ISpS. Furthermore, if c0 c1 c2 0, then the system is ISS. Remark 2.1 As noticed in the Fact, it follows from (6.21) that the x1 -system is ISpS. Moreover, the following ISpS property can be proved: V1
x1
t 1
jx1
t0 j; t
t0 1
kV2
x2 t0 ;t k 1
ku1t0 ;t jc1 ;
t t0 0
6:24
where 1 is a class KL-function. Similarly, (6.22) implies that the x2 -system is ISpS and enjoys the following IOpS property: V2
x1
t 2
jx2
t0 j; t
t0 2
kV1
x1 t0 ;t k 2
ku2t0 ;t jc2 ;
t t0 0
6:25
In order to invoke Theorem 2.4 to conclude the ISpS and IOpS properties for the feedback system, we shall require the small gain condition (6.19) to hold between the gain functions 1 and 2 . This obviously leads to a stronger gain condition than (6.23).
6.3
Adaptive passivation
The main purpose of this section is to extend passivation tools in [2] to a nonlinear system with unknown parameters described by: x '0
x '
x
0
x
xu
6:26
Then, we show that these adaptive passivation tools can be recursively used to design an adaptive controller for a class of interconnected nonlinear systems with unknown parameters. Our results can be regarded as a passivity interpretation of the popular adaptive backstepping with tuning functions algorithm as advocated in [26]. 6.3.1 De®nitions and basic results We will introduce notions of adaptive stabilizability and adaptive passivation for a system of the form (6.26). Our de®nition of adaptive stabilizability was
128
Adaptive nonlinear control: passivation and small gain techniques
motivated by a related notion in [26, p. 132] but is applicable to a broader class of systems ± see Section 6.3.2 below. First, we de®ne two useful properties. De®nition 3.1 (COCS- and UO-functions) general form x_ f
x u.
Consider a control system of the
(i) Assume a dynamic controller u
x; where _
x; . A continuous function y
x; is converging-output converging-state (COCS) for the closed loop system if all the bounded solutions
x
t;
t satisfy the following implication f
x
t;
t ! 0g
) fx
t ! 0g
6:27
n
(ii) A function V R ! R is said to be unboundedness observable (UO) if for any initial condition x
0 and any control input u 0; T ! Rm , with 0 < T 1, the following holds fjx
tj
t!T
!
1g
)
fV
x
t
t!T
!
1g
6:28
for some 0 < T T. Note that any proper function V
x is a UO-function for system x f
x; u, but the converse is not true. Similarly, a sucient condition for being a COCS-function is that is positive de®nite in x for each . Throughout the chapter, we use to denote the update estimate of the unknown parameter and restrict ourselves to the case where dim dim . In other words, overparametrization is avoided in that the number of parameter estimates equals to the number of unknown parameters. De®nition 3.2 (Adaptive stabilizability) The system (6.26) is said to be globally adaptively (quadratically) stabilizable if there exist a smooth UO-function
and an adaptive controller V
x; for each , an adaptive law
x;
u #
x; such that the time derivative of the augmented Lyapunov function candidate
V
x;
1
T 1
V
x;
6:29 2 satis®es
V
1
x;
6:30
is a COCS-function for the closed loop system (6.26). where 1
x; Remark 3.1 A system which satis®es this property is stabilizable in the sense that all the solutions of the closed loop system are bounded and that x
t tends to zero as t ! 1. Indeed, the boundedness property follows from (6.29) and
goes (6.30). In addition, by application of Barbalat's lemma [24], 1
x
t;
t to zero as t ! 1. Since 1 is a COCS-function, the claim follows readily. De®nition 3.3 (Adaptive passivation) The system (6.26) is said to be adaptively (quadratically) feedback passive (AFP) if there exist smooth functions
Adaptive Control Systems
129
V
x; 0, # and , with V
0; 0 8, and a C 0 function h such that the resulting system with adaptive feedback _
;
x;
v u #
x;
6:31
is passive with respect to input
v; and output h and the storage function V of the form (6.29). That is v _ T
V 2
x; h
x; ;
6:32 where 2 is a nonnegative function.
is a COCS-function for the zero-input closed loop system (6.26), If 2
x; then system (6.26) is said to be strongly AFP. The following result, which is an adaptive version of [2, Proposition 4.14], shows that the above notions of adaptive stabilizability and adaptive passivation are equivalent. Proposition 3.1 A system of the form (6.26) is strongly AFP with a UOfunction V for (6.26) if and only if it is globally adaptively stabilizable. Proof
The necessity is obvious. To prove the suciency, de®ne h by
0
x
x;
T 1
@V
x;
hT
x; ; @x
which completes the proof by Lemma 2.1. Let us point out that the problem of adaptive passivation cannot be trivially solved using the existing (known parameter) passivation tools. To this end, consider a scalar nonlinear system x_ x2 x u yx
6:33
where is an unknown constant parameter. If were known, system (6.33) would be made passive via the change of feedback law u x x2 x v. However, in the present case where is unavailable for feedback design, it is not clear how to ®nd a feedback transformation of the form (6.8) so that the resulting system is passive for every in R. In fact, it can be shown that no memoryless continuous statefeedback law of the form (6.8) achieves the passivation goal regardless of the value of . We now give an adaptive passivation result for cascaded nonlinear systems in the form of (6.26) but without unknown parameters in the input space. This result was motivated by Theorem 6.3 in which all parameters are known.
130
Adaptive nonlinear control: passivation and small gain techniques
!"orem 3.1
Consider a cascaded nonlinear system of the form _ f0
l X
f1i
; xyi
i1
6:34
x_ f
x f
x g
x
u g
x y h
x
where is a vector of unknown constant parameters. Suppose _ f0
is GAS at 0. It is also assumed that f f ; g; hg is strictly r C -passive with a positive de®nite and proper storage function (r 1. If there exists a matrix g1 of appropriate dimension such that
f
x g
xg1
x;
8 x 2 Rn
6:35
then system (6.34) is strongly adaptively feedback passive with a proper storage function V. Proof By a Converse Lyapunov Theorem [27], there exists a C1 Lyapunov function U which is positive de®nite and proper satisfying @U
f0
@
6:36
where is a positive de®nite function. Let W be a C r storage function associated with f f ; g; hg and consider the augmented Lyapunov function candidate
U
W
x 1
V
; x; 2
T 1
6:37
By hypotheses and Lemma 2.1, it follows that V_
l X @U i1
@
f1i
; xyi
@W
xf
x @x
yu
g1
x g
x
_ T 1
6:38
@W
xf
x is negative de®nite in x. @x For each
; x, denote the m l matrix c
; x as
Notice that
c
; x
f11
; xT
@U T @U T
; ; f1l
; xT
@ @
6:39
Adaptive Control Systems
131
Then, (6.38) yields V_
@W
x f
x yT u c
; x
g1
x g
x
@x
_ T 1
6:40
Thus, by choosing the following parameter update law and adaptive controller _
c
; x g1
x g
xT yT u
y
6:41
c
; x g1
x g
x
6:42
with a `®rst-sector third-sector' function, (6.40) implies @W
x f
x yT
y
6:43 @x
it follows from (6.43) that all the Since V is proper in its argument
; x; ,
solutions
t; x
t;
t of the closed loop system (6.34), (6.41) and (6.42) are well-de®ned and uniformly bounded on 0; 1. Furthermore, a direct application of LaSalle's invariance principle ensures
that all the trajectories
t; x
t;
t converge to the largest invariant set E
j
; x
0; 0g. Therefore, x
t tends to contained in the manifold f
; x; zero as t goes to 1. In other words, the cascade system (6.34) is globally adaptively stabilized by (6.41) and (6.42). Finally, Proposition 3.1 ends the proof of Theorem 3.1 V_
Remark 3.2 dim , then
It is of interest to note that, if rank fg
0
g1
0 g
0g l
j
t
t!1
j 0
Indeed, on the set E, we have g
0
g1
0 g
0
. So, E f
0; 0; g. implies that
t
t 0 which, in turn,
The following corollary is an immediate consequence of Theorem 3.1 where the -system in (6.34) is void. Corollary 3.1
Consider a linearly parametrized nonlinear system in the form x f
x f
x g
x
u g
x y h
x
6:44
If f f ; g; hg is strictly C 1 -passive with a positive de®nite and proper storage function, and if (6.35) holds, then system (6.44) is strongly adaptively C 1 feedback passive with a proper storage function V. 6.3.2 Recursive adaptive passivation In this section, we show that the adaptive passivation property can be
132
Adaptive nonlinear control: passivation and small gain techniques
propagated via adding a feedback passive system with linearly appearing parameters. This is indeed the design ingredient which was used in [19]. More precisely, consider a multi-input multi-output nonlinear system of the form (6.26) with linear parametrization: _ f10
f1
G10
G1
y
6:45
z_ q
z y y_ f20
x f2
x
G20
x G2
xu with x
T ; zT ; yT T , 2 Rn0 , z 2 Rn G20
x G2
x.
m
and y, u 2 Rm . Denote G2
x;
Proposition 3.2 If the -subsystem of (6.45) with y considered as input is AFP, if G2 is globally invertible for each , then the interconnected system (6.45) is also AFP. Furthermore, under the additional condition that the z-system is BIBS (bounded-input bounded-state) stable and is GAS at z 0 whenever
; y
0; 0, if the -system has a UO-function V1 and a COCS-function 1 associated with its AFP property, then the whole composite system (6.45) also possesses a UO-function V2 and a COCS-function 2 associated with its AFP property. Remark 3.3 Under the conditions of Proposition 3.2, it follows from Theorem 2.2 that the
z; y-system in (6.45) is feedback passive for every frozen and each . _ Proof Introduce the extra integrator 0 where 0 is a new input to be built recursively. By assumption, there exist a smooth positive semide®nite function
smooth functions #1 and 1 as well as a nonnegative function 1 and a V1
; , continuous function h1 such that the time derivative of the function
1
V 1 V1
; 2
T 1
satis®es V_ 1
hT
; ;
1
; 1
y 1
6:46
6:47
and 1 0 1
; .
where y y #1
; Letting h1
hT11 ; hT12 T , it follows from Lemma 2.1 that hT11
@V1
1
G1
;
;
G @
hT12
@V1
; @
T
1
6:48
is ane in . Then, there exist smooth functions h 11 which implies h11
; ; and h11 such that
h 11
;
h11
;
h11
; ;
6:49
Adaptive Control Systems
133
Consider the nonnegative functions
1 jy V2 V1
2 V 2 V 1 12 jy
2 1
j
6:50
2 #1
; j
6:51
In view of (6.47) and the de®nition of y and 1 , the time derivative of V 2 along _ the solutions of (6.45) and 0 satis®es @V1
T 1 1
hT
; ;
; V_ 2 1
; y 11 @
y T f20
x f2
x G2
x; u @#1
1 1 @
@#1 f10
f1
G10
G1
yj1 @
6:52
We wish to ®nd changes of feedback laws which are independent of
v u #2
x;
6:53
0 1 1 2 2 such that V 2 satis®es a dierential dissipation inequality like (6.47). To this end, set
2
T2 T2 y 2 2 2
6:55
where 2 , 2 and 2 2 Rml are de®ned by @#1 @#1
; G11
y; ;
; G1p
y 2 @ @ @#1 f1
@
6:54
6:56
2
x;
6:57
; G2p
x#2
x;
2
G21
x#2
x; ;
6:58
2 f2
x h11
;
Noticing that @V1
; @
T
1
1
@V1 T T
;
y 2 2 @
T
T2 T2 y @V1 T 1
;
2 @
6:59
134
Adaptive nonlinear control: passivation and small gain techniques
and that
V2
V1
x
y T
1
6:60
with (6.52), (6.53) and (6.55), simple computation yields
@V2
x;
T 1 2 V_ 2 1
; @
2
f20
x f2
x @#1
1 2 G2
x; v G2
x; # y T h 11
; @
@ T V1 @#1
f10
f1
G10
G1
y
2 2 @ @
T 1 2 y T G2
x; v
@V2
x; 1
; @
T
G2
x; #
2
x;
2 @ V1
;
6:61 y T 2
x; @
Observing that 2 depends on #2 , the following variable is introduced to split this dependence @ T V1
2 Rl
; @
With (6.58)
2
x;
2 G2
x; #
@ T V1
G2
x; #2
x;
; @
6:62
Since G2
x; is globally invertible, by choosing #2 as
G2
x; 1
2
x;
#2
x; and de®ning h2
hT21 ; hT22 T
6:63
by
y; h21 GT2
x;
h22
@ T V2
1
x;
@
we obtain V 2
y
hT
x; ;
2
x; 2
v 2
6:64
6:65
j yj2 . The ®rst statement of Proposition 3.2 was proved. with 2 1
; The second part of Proposition 3.2 follows readily from our construction and the main result of Sontag [43]. On the basis of Corollary 3.1 and Proposition 3.1, a repeated application of Proposition 3.2 yields the following result on adaptive backstepping stabilization.
Adaptive Control Systems
Corollary 3.2 [26]
135
Any system in strict-feedback form x_ i xi1 i
x1 xi ; x_ n u n
x1 ; ; xn
1in
1
6:66
is globally adaptively (quadratically) stabilizable. 6.3.3 Examples and extensions We close this section by illustrating our adaptive passivation algorithm with the help of cascade-interconnected controlled Dung equations. Possible extensions to the output-feedback case and nonlinear parametrization are brie¯y discussed via two elementary examples. 6.3.3.1 Controlled Cung equations
Consider an interconnected system which is composed of two (modi®ed) Dung equations in controlled form, i.e. x1 1 x_ 1 1 x1 2 x31 u1 x2 2 x_ 2 3 x2 4 x32 u2
6:67
where 1 ; 2 > 0 are known parameters,
1 ; 2 ; 3 :4 is a fourth-order vector of unknown constant parameters and u2 is the control input. The interconnection constraint is given by u1 2 x2 x_ 2
6:68
Denoting z1 x1 and z2 x2 , the coupled Dung equations (6.67) can be transformed into the following state-space model z_ 1
1 z 1 y1 ;
y_ 1
1 z1
2 z31 y2 ;
z_ 2
2 z 2 y2 ;
y_ 2
3 z2
4 z32 u2
6:69
Obviously, the
z1 ; y1 -system in (6.69) is AFP by means of the change of parameter update law and adaptive controller _ 1 _ 2 y2
1 y1 z1 11 ;
1 > 0
2 y1 z31
12 ; 2 > 0 y1 1 z1 2 z3 y 2 1
In addition, V1 12 y21 is a UO-function for the
z1 ; y1 -system which satis®es the dierential dissipation equality 1 1 2 2 ;
y2 ; 1 T
6:70 V_ 1 y21
y1 ; 1 2 with V 1 V1
y1
1
1 21
1 2
1
2 22
2 2 .
136
Adaptive nonlinear control: passivation and small gain techniques
It is easy to check that the conditions of Proposition 3.2 hold. A direct application of our adaptive passivation method in the proof of Proposition 3.2 gives our passivity-aimed adaptive stabilizer for system (6.69), or the original system (6.67): _ 1
1
2y1 y2
1 z1
2 z31 z1
_ 2
2
2y1 y2
1 z1
2 z31 z31
_ 3
3
y1 y2
1 z1
2 z31 z2 ;
3 > 0
_ 4
4
y1 y2
1 z1
2 z31 z22 ;
4 > 0
u
6.3.3.2
2y1
6:71
_ _ 2y2 3 z2 4 z32 1 z1 2 z31
1 3 2 z31
1 z1 y1
Adaptive output feedback passivation
The adaptive passivation results presented in the previous sections rely on fullstate feedback (6.31). In many practical situations, we often face systems whose state variables are not accessible by the designer except the information of the measured outputs. Unlike the state feedback case, the minimum-phase and the relative degree-one conditions are not sucient to achieve adaptive passivation if only the output feedback is allowed. This is the case even in the context of (nonadaptive) output feedback passivation, as demonstrated in [40] using the following example: z_
z3
_ z u
6:72
y It was shown in [40, p. 67] that any linear output feedback u ky v, with k > 0, cannot render the system (6.72) passive. In fact, Byrnes and Isidori [1] proved that the system (6.72) is not stabilizable under any C 1 output feedback law. As a consequence, this system is not feedback passive via any C1 output feedback law though it is feedback passive via a C1 state feedback law. However, system (6.72) can be made passive via the C0 output-feedback given by 1 3 u ky3 v; k > 4=3
6:73 2 Indeed, consider the quadratic storage function V 12 z2 12 2
6:74
Adaptive Control Systems
137
Forming the derivative of V with respect to the solutions of (6.72), using Young's inequality [8] gives V_
z4 2z
1
z4
4
k3 yv
k
3 243 13
43 yv
6:75
27 < < 1. 16k3 Therefore, system (6.72) in closed loop with output feedback (6.73) is (state) strictly passive. As seen from this example, the output feedback passivation issue is more involved and requires additional conditions on the system or nonsmooth feedback strategy. Thus, it is not surprising that the problem of adaptive output feedback passivation is also complex and solving it needs extra conditions in addition to minimum phaseness and relative-degree one. As an illustration, let us consider a nonlinearly parametrized system with outputdependent nonlinearity: z_ z3 where
_ z u '
y;
6:76
y where is a vector of unknown constant parameters. Assume that the nonlinear function ' checks the following concavity-like condition. (C) For any y and any pair of parameters
1 ; 2 , we have y'
y; 2
y'
y; 1 y
@'
y; 2
2 @
1
6:77
Non-trivial examples of ' verifying such a condition include all linear parametrization (i.e. '
y; '1
y) and some nonlinearly parametrized functions like '
y; '2
y
'3
y '1
y where y'2
y 0 for all y 2 R. Consider the augmented storage function V V
z; 12
T 1
where is an update parameter to be preÂcised later. By virtue of (6.73) and (6.75), we have 3 4 V
1 z k 4=3 y
v '
y;
24=3 1=3
6:78
T 1
6:79
138
Adaptive nonlinear control: passivation and small gain techniques
Letting v
v '
y
_
T y
y
it follows from the condition (C) and (6.79) that 3 _ 4 v
4=3 y V
1 z k 24=3 1=3
6:80
T 1
6:81
In other words, the system (6.76) is made passive via adaptive output-feedback law (6.73)±(6.80). In particular, the zero-input closed-loop system (i.e.
0; 0; and, furthermore, the
v;
0; 0) is globally stable at
z; ; trajectories
z
t;
t go to zero as t goes to 1.
6.4
Small gain-based adaptive control
Up to now, we have considered nonlinear systems with parametric uncertainty. The synthesis of global adaptive controllers was approached from an input/ output viewpoint using passivation±a notion introduced in the recent literature of nonlinear feedback stabilization. The purpose of this section is to address the global adaptive control problem for a broader class of nonlinear systems with various uncertainties including unknown parameters, time-varying and nonlinear disturbance and unmodelled dynamics. Now, instead of passivation tools, we will invoke nonlinear small gain techniques which were developed in our recent papers [21, 20, 16], see references cited therein for other applications.
6.4.1 Class of uncertain systems The class of uncertain nonlinear systems to be controlled in this section is described by z_ q
t; z; x1 x_ i xi1 T 'i
x1 ; ; xi i
x; z; u; t; x_ n u T 'n
x1 ; ; xn n
x; z; u; t
1in
1
6:82
y x1 where u in R is the control input, y in R is the output, x
x1 ; ; xn is the measured portion of the state while z in Rn0 is the unmeasured portion of the state. in Rl is a vector of unknown constant parameters. It is assumed that the
i 's and q are unknown Lipschitz continuous functions but the 'i 's are known smooth functions which are zero at zero.
Adaptive Control Systems
139
The following assumptions are made about the class of systems (6.82). (A1) For each 1 i n, there exist an unknown positive constant p i and two known nonnegative smooth functions i1 , i2 such that, for all
z x u t j i
x z u tj p i
i1
j
x1 xi j
p i
i2
jzj
6:83
Without loss of generality, assume that i2
0 0. (A2) The z-system with input x1 has an ISpS-Lyapunov function V0 , that is, there exists a smooth positive de®nite and proper function V0
z such that @V0
zq
t; z; x1 @z
0
jzj 0
jx1 j d0
8
z; x1
6:84
where 0 and 0 are class K1 -functions and d0 is a nonnegative constant. The nominal model of (6.82) without unmeasured z-dynamics and external disturbances i was referred to as a parametric-strict-feedback system in [26] and has been extensively studied by various authors±see the texts [26, 32] and references cited therein. The robustness analysis has also been developed to a perturbed form of the parametric-strict-feedback system in recent years [48, 22, 31, 35, 51]. Our class of uncertain systems allows the presence of more uncertainties and recovers the uncertain nonlinear systems considered previously within the context of global adaptive control. The theory developed in this section presupposes the knowledge of partial xstate information and the virtual control coecients. Extensions to the cases of output feedback and unknown virtual control coecients are possible at the expense of more involved synthesis and analysis ± see, for instance, [17, 18]. An illustration is given in subsection 6.4.4 via a simple pendulum example.
6.4.2
Adaptive controller design
6.4.2.1 Initialization
We begin with the simple x1 -subsystem of (6.82), i.e. x_ 1 x2 T '1
x1 1
x; z; u; t
6:85
where x2 is considered as a virtual control input and z as a disturbance input. Consider the Lyapunov function candidate V1 12
x21 12
T 1
1
p 2
p2
6:86
where > 0; > 0 are two adaptation gains, is a smooth class-K1 function to be chosen later, p fp?i ; p?i 2 j1 i ng is an unknown constant and the
Adaptive nonlinear control: passivation and small gain techniques
140
p are introduced to diminish the eects of parametric time-varying variables , uncertainties. With the help of Assumption (A1), the time derivative of V1 along the solutions of (6.82) satis®es: V_ 1 0
x21 x1 x2 T '1
x1 p 1 0
x21 jx1 j 11
jx1 j 12
jzj _ 1 p T 1
pp _
6:87
where 0
x21 is the value of the derivative of at x21 . In the sequel, is chosen such that 0 is nonzero over R . R1 Since 11 is a smooth function and 11
jx1 j 11
0 jx1 j 0 011
sjx1 j ds, given any "1 > 0, there exists a smooth nonnegative function 1 such that p 1 0
x21 jx1 j
11
jx1 j
p0
x21 x21 1
x1 "1
11
0
2
;
8x1 2 R
6:88
By completing the squares, (6.87) and (6.88) yield V_ 1 0
x21 x1
x2 T '1
x1 px1 1
x1 p 14 x1 0
x21
1 p
pp _
2 12
jzj
"1
11
0
_ T 1
2
6:89
De®ne 0
x21 x1 '1
x1
6:90
$1
p p x21
1
x1 14 0
x21 0
x21
6:91
#1
x1 1
x21
1
w2 x2
T '1
x1
p
x1 1
x1 14 x1 0
x21
p #1
x1 ; ;
6:92
6:93
where ; p > 0 are design parameters, 1 is a smooth and nondecreasing function satisfying that 1
0 > 0. Consequently, it follows from (6.89) that V_ 1
0 x21 1
x21 0 x1 w2 1 p
p
p _
$1
T
2 12
jzj
"1
p p
11
0
2
p
p
_ T 1
1
6:94
It is shown in the next subsection that a similar inequality to (6.94) holds for each
x1 ; ; xi -subsystem of (6.82), with i 2; ; n. 6.4.2.2 Recursive steps Assume that, for a given 1 k < n, we have established the following property (6.95) for the
x1 ; ; xk -subsystem of system (6.82). That is, for each
Adaptive Control Systems
141
1 i k, there exists a proper function Vi whose time derivative along the solutions of (6.82) satis®es i X
cj ijw2j wi wi1
T p
p
0 x21
1
x21 i1
V_ i
p
p
j2
i X
T
i X
1
j2
j
2
i j12
jzj
@#j 1 _
wj @
i X
j"i
1
p i
j1
i j11
0
p
i X j2
@#j 1 _ wj
p $i @ p
2
6:95
j1
j1
In (6.95), "j > 0
1 j i are arbitrary, cj > n j
2 j i are design parameters, #j
1 j i, i and $i are smooth functions and the variables wj 's are de®ned by w1 x1 0
x21 ;
p ; #j 1
x1 ; ; xj 1 ; ;
wj xj
2j i1
6:96
p 0 for each pair of
;
p and all It is further assumed that #i
0; ; 0; ; 1 i k. The above property was established in the preceding subsection with k 1. In the sequel, we prove that (6.95) holds for i k 1. Consider the Lyapunov function candidate
p 1 w2 Vk1 Vk
x1 ; w2 ; ; wk ; ; 2 k1
6:97
In view of (6.95), dierentiating Vk1 along the solutions of system (6.82) gives V k1 0 x21
1
x21 k1
k X
cj k jw2j wk wk1
T p
p p
p
j2
T
1
k X j2
k X @#j 1
1 @#j 1
p p wj
k wj
p $k @ p
@
j2
k X @#k @#k @#k wk1 xk2 T 'k1 k1
xj1 T 'j j p
@xj @ p
@
j1
k X j1
j
2
k j12
jzj
k X
j"k
j1
k j11
0
2
6:98
j1
Recalling that p fp?i ; p?i 2 j1 i ng, by virtue of assumption (A1), we
142
Adaptive nonlinear control: passivation and small gain techniques
have 2 k k X X
k
k 2 1 1
j pwk1 4 4 wk1 k1
xj
xj j1 j1 jwk1 j p k1
k11
j
x1 xk1 j
k X
k j1
j
x1 xj j p j
xj j1
k1 X
2 j2
jzj
j1
6:99
From (6.93) and (6.96), it is seen that
w1 ; ; wk ; wk1
0; ; 0; 0 if and only if
x1 ; ; xk ; xk1
0; ; 0; 0. Recall that 0
x21 6 0 by selection. With this observation in hand, given any "k1 > 0, lengthy but simple calculations imply the existence of a smooth nonnegative function k1 such that k X @#k pj jwk1 j p k1
k11
j
x1 ; ; xk1 j j1
j
x1 ; ; xj j @xj j1
p x2 0
x2 pw2k1 k1
w1 ; w2 ; ; wk1 ; ; 1 1
k X
w2j
j2
k1 X
"j
j1
0
2
6:100
j1
Combining (6.98), (6.99) and (6.100), we obtain 0 x21
1
x21
V_ k1
k X
k
cj
1 jw2j
k
T
p p
p
p
j2
T
1
k X
wj
j2
@#j 1 _
1
p
k @
wk1 xk2 wk T 'k1 pwk1
k1 X j1
j
k X @#k j1
k 1 1X @#k 2
k1
4 4 j1 @xj
k j22
jzj
2
k1 X
j"k
@xj
p
k X j2
wj
@#j 1 _
p $k @ p
xj1 T 'j
@#k _
@
j2
k j21
0
@#k _ p
@ p
2
6:101
j1
Inspired by the tuning functions method proposed in [26, Chapter 4] for
Adaptive Control Systems
143
parametric-strict-feedback systems without unmodelled dynamics, introduce the notation k X
k k1 k 'k1 j wk1
6:102
xj j1 $k1 $k
1 4
1 4
wk
k X @#k j1
T
k X j2
wk1 p
@#j 1 wj @
k X j2
k1 w2k1
@xj
j1
ck1 w2k1
#k1
2 k X @#k
xj1
@xj
@#j 1 wj @ p
6:103
k X @#k
'k1
j1
1 4
1 4
@xj
2 k X @#k @xj
j1
@#k @#k $k1 wk2 k1
@ p
@
'j
k1
6:104
p #k1
x1 ; ; xk1 ; ;
wk2 xk2
6:105
Therefore, inequality (6.95) follows readily after equalities (6.102) to (6.105) are substituted into (6.101). By induction, at the last step where k n in (6.95), if we choose the following parameter update laws and adaptive controller _
p ; n
x1 ; ; xn ; ;
p p _ $n
x1 ; ; xn ; ;
6:106
p u #n
x1 ; ; xn ; ;
6:107
the time derivative of the augmented Lyapunov function Vn 12
y2
n X
xi
#i 1 2 12
T 1
n iw2i
i2
1
p 2
p2
6:108
satis®es 0
x21 x21
1
x21 n 1
V_ n
n X
ci
T
p
p
p
p
i2
n X i1
i
2
n i12
jzj
n X
i"n
i1
n i11
0
2
6:109
i1
As a major dierence with most common adaptive backstepping design
144
Adaptive nonlinear control: passivation and small gain techniques
procedures [26, 32], because of the presence of dynamic uncertainties z, we are unable to conclude any signi®cant stability property from the inequality (6.109). Another step is needed to robustify the obtained adaptive backstepping controllers (6.106) and (6.107). 6.4.2.3 Small gain design step The above design steps were devoted to the x-subsystem of (6.82) with z considered as the disturbance input. The eect of unmeasured z-dynamics has not been taken into account in the synthesis of adaptive controllers (6.106) and (6.107). The goal of this section is to specify a subclass of adaptive controllers in the form of (6.106), (6.107) so that the overall closed loop system is Lagrange stable. Furthermore, the output y can be driven to a small vicinity of the origin if the control design parameters are chosen appropriately. First of all, the design function 1 as introduced in subsection 6.4.2.1 is selected to satisfy 0
x21 x21
1
x21 n 1 c1
x21
6:110 with c1 > 0 a design parameter. Such a smooth function always exists because 0
x21 6 0 for any x1 . Then, let 1 be a smooth class-K1 function which satis®es the inequality n X
i
n i12
jzj
2
1
jzj2
6:111
i1
Noticing that
T
p
p (6.106) yields:
p
p
T 1
2
1 p p
p p2 p2 2 2 V n
"1
2 jj 2
n i;
n 2 p 2 X jj p i"n 2 2 i1
6:114
; p 2 i ng
1 i1
n i11
0
6:112
6:113
cVn 1
jzj2 "1
where c f2c1 ; 2
ci
2
6:115
6:116
Let v and v be two class-K1 functions such that v
jzj V0
z v
jzj
6:117
Given any 0 < "2 < c, (6.114) ensures that V n
"2 Vn
6:118
Adaptive Control Systems
145
whenever Vn
2 c
2"1 1
v
V0
z ; 2 c 2 1
2
6:119
Return to the z-subsystem. According to assumption (A2), we have @ V0
zq
z; x1 "3 0
jzj
1 "3 0
jx1 j "3 d0
6:120 @z where V0 "3 V0 , "3 > 0 is arbitrary and is a class-K1 function to be determined later. Given any 0 < "4 < 1, we obtain V 0
as long as V0
2
1
"3 v 0
1
4
"3 "4 0
jzj
1
0
jx1 j; 3 v 0
6:121 1
2d0 1 4
6:122
To check the small gain condition (6.23) in Theorem 2.5, we select any classK1 function such that s 1 "4 c 2 0 v 1 v s ; 8s > 0
6:123 1 1
s < 2 4 Finally, to invoke the Small Gain Theorem 2.5, it is sucient to choose the function appropriately so that
1
0
jx1 j 12
x21 "5
6:124
where "5 > 0 is arbitrary. In other words,
1
p "5 0
jx1 j Vn
x1 ; x2 ; ; xn ; ;
6:125
Clearly, such a choice of the smooth function is always possible. Consequently, V 0 "3 "4 0
jzj
6:126 as long as 2 2 2d0 1 1 1 V0 "3 v 0
2Vn ; "3 v 0
2"5 ; "3 v 0 1 4 1 4 1 "4
6:127
Under the above choice of the design functions and 1 , the stability properties of the closed loop system (6.82), (6.106) and (6.107) will be analysed in the next subsection.
146
Adaptive nonlinear control: passivation and small gain techniques
6.4.3 Stability analysis If we apply the above combined backstepping and small-gain approach to the plant (6.82), the stability properties of the resulting closed loop plant (6.82), (6.106) and (6.107) are summarized in the following theorem. !"orem 4.1 Under Assumptions (A1) and (A2), the solutions of the closed loop system are uniformly bounded. In addition, if a bound on the unknown parameters p i is available for controller design, the output y can be driven to an arbitrarily small interval around the origin by appropriate choice of the design parameters. Proof Letting and p p p, it follows that Vn is a positive de®nite p. Also, _ _ and p_ p _ . Decompose the and proper function in
x1 xn ; closed-loop system (6.82), (6.106) and (6.107) into two interconnected sub p-subsystem and the other is the z-subsystem. systems, one is the
x1 ; ; xn ; ; We will employ the Small Gain Theorem 2.5 to conclude the proof. p-subsystem. From (6.118) and (6.119), it Consider ®rst the
x1 ; ; xn ; ; follows that a gain for this ISpS system with input V0 and output Vn is given by 1
s
2 c
2 1 1 v 1 s 2 "3
6:128
Similarly, with the help of (6.126) and (6.127), a gain for the ISpS z-subsystem with input Vn and output V0 is given by 2
s 3 v 0 1
2 1
"4
2s
6:129
As it can be directly checked, with the choice of as in (6.123), the small gain condition (6.23) as stated in Theorem 2.5 is satis®ed between 1 and 2 . Hence, a direct application of Theorem 2.5 concludes that the solutions of the interconnected system are uniformly bounded. The second statement of Theorem 4.1 can be proved by noticing that the drift constants in (6.119) and (6.127) can be made arbitrarily small. Remark 4.1 It is of interest to note that the adaptive regulation method presented in this section can be easily extended to the tracking case. Roughly
i speaking, given a desired reference signal yr
t whose derivatives yr
t of order up to n are bounded, we can design an adaptive state feedback controller so that the system output y
t remains near the reference trajectory yr
t after a considerable period of time. Remark 4.2 Our control design procedure can be applied mutatis mutandis to a broader class of block-strict-feedback systems [26] with nonlinear
Adaptive Control Systems
147
unmodelled dynamics z_ q
t z x1 x_ i xi1 T i
x1 xi ; 1 ; ; i xi
x; ; z; u; t _i i;0
x1 ; ; xi ; 1 ; ; i T i
x1 ; ; xi ; 1 ; ; i i
x; ; z; u; t;
6:130
1in
where xn1 u, x
x1 ; ; xn T and
1 ; ; n T . Assume that all i dynamics are measured and satisfy a BIBS stability property when
x1 ; ; xi ; 1 ; ; i 1 ; z is considered as the input. Similar conditions to (6.83) are required on the disturbances xi and i . 6.4.4 Examples and discussions We demonstrate the eectiveness of our robust adaptive control algorithm by means of a simple pendulum with external disturbances. Along the way, we show that our combined backstepping and small gain control design procedure can be extended to cover systems with unknown virtual control coecients. Moreover, we shall see that the consideration of dynamic uncertainties occurring in our class of systems (6.82) becomes very natural when the output feedback control problem is addressed. Then, in the subsection 6.4.4.2, we compare the above adaptive design method with the dynamic normalization-based adaptive scheme proposed in our recent contribution [18] via a second order nonlinear system. 6.4.4.1 Pendulum example The following simple pendulum model has been used to illustrate several nonlinear feedback designs (see, e.g., [3,24]): ml
mg
1 kl u 0
t l
6:131
where u 2 R is the torque applied to the pendulum, 2 R is the anticlockwise angle between the vertical axis through the pivot point and the rod, g is the acceleration due to gravity, and the constants k, l and m denote a coecient of friction, the length of the rod and the mass of the bob, respectively. 0
t is a time-varying disturbance such that j 0
tj a0 for all t 0. It is assumed that these constants k, l, m and a0 are unknown and that the angular velocity is not measured. Using the adaptive regulation algorithm proposed in subsection 6.4.3, we want to design an adaptive controller using angle-only so that the pendulum is kept around any angle < 0 .
148
Adaptive nonlinear control: passivation and small gain techniques
We ®rst introduce the following coordinates 1 ml 2
0 k 2 ml 2 _
m
0
6:132
_
0 ; 0 into the origin
1 ; 2
0; 0. to transform the target point
; It is easy to check that the pendulum model (6.131) is written in -coordinates as _1 2
k 1 m
6:133
_2 u
1 mgl 0 2 1 l 0
t ml
6:134
Since the parameters k, l and m are unknown and the angular velocity is unmeasured, the state
1 ; 2 of the transformed system (6.133)±(6.134) is therefore not available for controller design. We try to overcome this burden with the help of the `Separation Principle' for output-feedback nonlinear systems used in recent work (see, e.g., [23, 26, 32, 36]). Here, an observer-like dynamic system is introduced as follows 1 2 l1
2 u l2
0 0
1
6:135
1
where l1 and l2 are design parameters. Denote the error dynamics e as e
1
1 ; 2
2 T
6:136
Noticing (6.132), the e-dynamics satisfy 3 2 l1 k l1 1 7 6 l1 1 ml 2 m 7 6 e e6 7 5 4 l2 1 l2 0 |{z} mgl
t l l2 1 1 0 0 2 2 ml ml A
6:137
For the purpose of control law design, let us choose a pair of design parameters l1 and l2 so that A is an asymptotically stable matrix. Letting x1 0 , x2 2 and z e=a with a jl1 ml 2 l1 kl 2 j; jl2 ml 2 l2 j; mgl; la0
6:138
Adaptive Control Systems
149
we establish the following system to be used for controller design: z_ Az x 1
x1
l1 ml 2 x1
l2 ml 2
1 x2 ml 2
x 2 u l2
1
l2 a
l1
kl 2 a
mgla
x1 0
la 0
k 1 x1 2 az2 m ml
6:139
ml 2 x1 l2 az1
y x1 1 , referred to as a `virtual control coecient' ml 2 [26], occurs before x2 , this system (6.139) is not really in the form (6.82). Nonetheless, we show in the sequel that our control design procedure in subsection 6.4.3 can be easily adapted to this situation. Let P > 0 be the solution of the Lyapunov equation Since the unknown coecient
PA AT P
6:140
2I
Then, it is directly checked that along the solutions of the z-system in (6.139) the time derivation of V0 zT Pz satis®es V 0
Step 1:
jzj2 3
Px21 8
P
6:141
Instead of (6.86), consider the proper function V1
1 2 1 p x
2ml 2 1 2
p2
6:142
where 1 1 k 2 a2 p a2 ; 2 ; 2 4 ; 2 ; 2 4 ml m l m m l
6:143
It is important to note that we have not introduced the update parameter 1 for the unknown but negative parameter k=m because the term
k=mx1 is stabilizing in the x1 -subsystem of (6.139). The time derivative of V1 along the solutions of (6.139) yields: V 1
1 x21 x1 w2
p
p
1 p p
p
p
p
$1 12 z22
6:144
150
Adaptive nonlinear control: passivation and small gain techniques
where 1 > 0 is a design parameter, $1 and w2 are de®ned by p p
6:145
#1
1 x1
6:146
w2 x2 Step 2:
x21 ; 2 x1 p ; 2
$1
#1
x1 ; p
6:147
Consider the proper function V2 V1
1 2 2 w2
6:148
Then, with (6.144), the time derivative of V2 along the solutions of (6.139) satis®es V_ 2
1 p p
p _ $1 12 z22 1 x21 x1 w2 p
p p
p
p
1 k 1 x1 2 _ x1 2 az2 p
w2 ul2
1 ml x1 l2 az1 1 x2 2 ml 2 m ml 2
6:149
From the de®nition of x2 in (6.147) and p as in (6.143), we ensure that 4 p 1 p p
x 2 p 1 1
6:150 w22 14 x21 w2 1 2 ml 2 2 2 With the choice of p as in (6.143), by completing the squares, it follows from (6.149) and (6.150) that V_ 2
1 p p
p _ $1 jzj2 p
p
p p
p
x1 l2 _ w2 u x1 p p 1 1
1 2
1 4 2 w2 2 2 2 4 2
1
1x21
p
p
6:151
Thus, setting ! 2 4 2
l p p p 2 w22 1 p _ $1 1 1 1 4 2 2 2 u
2 w2
x1
x1 p _ 2
l2 p p
p
p 1 1
1 2
1 4 2 w2 4 2 2 2
6:152
Adaptive Control Systems
151
and substituting these de®nitions into (6.151), we establish V_ 2
1
1x21
2 w22
p
p 2
p2
p 2 p jzj2 2
6:153
with 2 > 0 a design parameter. In the present situation, it is easy to compute a storage function for the whole closed loop system from the above dierential inequalities (6.141) and (6.153). So, instead of pursuing the small gain design step as in subsection 6.4.2.3 where no storage function for the total system was given, we give such a storage function for the closed loop pendulum system. Indeed, consider the composite storage function V V0
z 12 V2
x1 ; x2 ; p
6:154
Clearly, from (6.141) and (6.153), it follows V_
0:5
1
1x21
0:52 w22
p
p 4
In other words V_ with c f
1
cV
p2
0:5jzj2
p 2 p 4
p 2 p 4
6:155
6:156
1ml 2 ; 2 ; 0:5p ; 0:5
P 1 g
Finally, from (6.156), it is seen that all the solutions of the closed loop system are bounded. In particular, the angle eventually stays arbitrarily close to the given angle 0 if an a priori bound on the system parameters m, l are known and the design parameters 1 , 2 , p and are chosen appropriately. 6.4.4.2 Robusti®cation via dynamic normalization It should be noted that an alternative adaptive control design was recently proposed in [17, 18] for a similar class of uncertain systems (6.82). The adaptive strategy in [17, 18] is a nonlinear generalization of the well-known dynamic normalization technique in the adaptive linear control literature [13] in that a dynamic signal was introduced to inform about the size of unmodelled dynamics. The adaptive nonlinear control design presented in this chapter yields a lower order adaptive controller than in [17, 18]. Nevertheless, due to the worse-case nature of this design, the consequence is that the present adaptive scheme may yield a conservative adaptive control law for some systems with parametric and dynamic uncertainties. Therefore, a co-ordinated design which exploits the advantages and avoids the disadvantages of these two adaptive control approaches is certainly desirable and this is left for future investigation.
152
Adaptive nonlinear control: passivation and small gain techniques
As an illustration of this important point, let us compare the two methods with the following simple example: z_
zx
6:157
x_ u x z2
6:158
where z is unmeasured and is unknown. Let us start with Method I: robust adaptive control approach without dynamic normalization as proposed in this chapter. Considering the z-subsystem with input x, assumption (A2) holds with V0 z2 whose time derivative satis®es jzj2 x2
V_ 0
6:159
In order to apply the Small Gain Theorem 2.5 we show that the x-system can be made ISpS (input-to-state practically stable) via an adaptive controller. An ISpS-Lyapunov function is obtained for the augmented system. To this purpose, consider the function W 12
x2
1
2
2
6:160
where > 0 and is a smooth function of class K1 . A direct computation implies: 1 _ W_ x0 u x z2
1 x0 z4 1
x0 u x 4
6:161 _
x2 0
where 0 stands for the derivative of . By choosing the adaptive law and adaptive controller _
x2 0
x2
u
x
x2
x
1 0 2 4 x
x
6:162
6:163
6:164
where > 0 and
> 0 is a smooth nondecreasing function, it holds: W_ Select so that Then (6.165) gives: with f2; g.
x2 0
x2
1
2
2 z4 12 2
x2 0
x2
x2
x2 W_
W z4
2 2
6:165
6:166
6:167
Adaptive Control Systems
153
_ _
Hence, letting ! and noticing ! , (6.167) implies that W is an ! ISpS-Lyapunov function for the x-system augmented with the -system when z is considered as the input. To complete our small gain argument, we need to choose an appropriate function so that a small gain condition holds. Following the small gain design step developed in subsection 6.4.2.3, we need to pick a function such that 2x4
x2 W 4 2
6:168
A choice of such a function to meet (6.168) is: 8
s s2
6:169
This leads to the following choice for
x2 12
6:170
and therefore to the following controller _ u
16 4 x 4 3
x x x
6:171
6:172
In view of (6.159) and (6.167), a direct application of the Small Gain Theorem
2.5 concludes that all the solutions
x
t; z
t;
t are bounded over 0; 1. In the sequel, we concentrate on the Method II: robust adaptive control approach with dynamic normalization as advocated in [17,18]. To derive an adaptive regulator on the basis of the adaptive control algorithm in [17], [18], we notice that, thanks to (6.159), a dynamic signal r
t is given by: r_ 0:8r x2 ; r
0 > 0
6:173 The role of this signal r is to dominate V0
z ± the output of the unmodelled eects ± in ®nite time. More precisely, there exist a ®nite T o > 0 and a nonnegative time function D
t such that D
t 0 for all t T o and V0
z
t r
t D
t;
8t 0
6:174
1 r 0
6:175
Consider the function V 12 x2
1
2
2
where 0 > 0. A direct application of the adaptive scheme in [18] yields the
Adaptive nonlinear control: passivation and small gain techniques
Input u
Output x
Input u
Output x
154
Figure 6.1 Method II with r
t versus Method I without r
t: the solid lines refer to Method II while the dashed lines to Method I
60
Parameter estimate
50
40
30
20
10
0 0
1
2
3 secs
4
5
6
Figure 6.2 Method II with r
t versus Method I without r
t: the solid lines refer to Method II while the dashed lines to Method I
Adaptive Control Systems
155
following adaptive regulator: _ x2 ; > 0 5 1 u x 4 0
6:176
x
0 xr 1:6
6:177
With such a choice, the time derivative of V satis®es: V_
x2
0:4 r 0
6:178
Therefore, all solutions x
t, r
t and z
t converge to zero as t goes to 1. Note that the adaptive controller (6.177) contains the dynamic signal r which is a ®ltered version of x2 while in the adaptive controller (6.172), we have directly x2 . But, more interestingly, the adaptation law (6.176) is in x2 whereas (6.171) is in x4 . As seen in our simulation (see Figures 6.1 and 6.2), for larger initial condition for x, this results in a larger estimate and consequently a larger control u. Note also that, with the dynamic normalization approach II, the output x
t is driven to +0.5% in two seconds. For simulation use, take 0:1 and design parameters 3 and r
0 0 1. The simulations in Figures 6.1 and 6.2 are based on the following choice of initial conditions: z
0 x
0 5;
0:5
0
Summarizing the above, though conservative in some situations, the adaptive nonlinear control design without dynamic normalization presented in this chapter requires less information on unmodelled dynamics and gives simple adaptive control laws. As seen in Example (6.157), the robusti®cation scheme using dynamic normalization may yield better performance at the price of requiring more information on unmodelled dynamics and a more complex controller design procedure. A robust adaptive control design which has the best features of these approaches deserves further study.
6.5
Conclusions
We have revisited the problem of global adaptive nonlinear state-feedback control for a class of block-cascaded nonlinear systems with unknown parameters. It has been shown that adaptive passivation represents an important tool for the systematic design of adaptive nonlinear controllers. Early Lyapunovtype adaptive controllers for systems in parametric-strict-feedback form can be reobtained via an alternative simpler path. While passivation is well suited for systems with a feedback structure, small gain arguments have proved to be more appropriate for systems with unstructured disturbances. From practical
156
Adaptive nonlinear control: passivation and small gain techniques
considerations, a new class of uncertain nonlinear systems with unmodelled dynamics has been considered in the second part of this chapter. A novel recursive robust adaptive control method by means of backstepping and small gain techniques was proposed to generate a new class of adaptive nonlinear controllers with robustness to nonlinear unmodelled dynamics. It should be mentioned that passivity and small gain ideas are naturally complementary in stability theory [5]. However, this idea has not been used in nonlinear control design. We hope that the passivation and small gain frameworks presented in this chapter show a possible avenue to approach this goal. Acknowledgements. This work was supported by the Australian Research Council Large Grant Ref. No. A49530078. We are very grateful to Laurent Praly for helpful discussions that led to the development of the result in subsection 6.4.4.2.
References [1] Byrnes, C. I. and Isidori, A. (1991) `Asymptotic Stabilization of Minimum-Phase Nonlinear Systems', IEEE Trans. Automat. Control, 36, 1122±1137. [2] Byrnes, C. I., Isidori, A. and Willems, J.C. (1991) `Passivity, Feedback Equivalence, and the Global Stabilization of Minimum Phase Nonlinear Systems', IEEE Trans. Automat. Control, 36, 1228±1240. [3] Corless, M. J. and Leitmann, G. (1981) `Continuous State Feedback Guaranteeing Uniform Ultimate Boundedness for Uncertain Dynamic Systems', IEEE Trans. Automat. Control, 26, 1139±1144. [4] Coron, J. -M., Praly, L. and Teel, A. (1995) `Feedback Stabilization of Nonlinear Systems: Sucient Conditions and Lyapunov and Input-Output Techniques'. In: Trends in Control (A. Isidori, ed.) 293±348, Springer. [5] Desoer, C. A. and Vidyasagar, M. (1975) `Feedback Systems: Input-Output Properties'. New York: Academic Press. [6] Fradkov, A. L., Hill, D. J., Jiang, Z. P. and Seron, M. M. (1995) `Feedback Passi®cation of Interconnected Systems', Prep. IFAC NOLCOS'95, 660±665, Tahoe City, California. [7] Hahn, W. (1967) Stability of Motion. Springer-Verlag. [8] Hardy, G., Littlewood, J. E. and Polya, G. (1989) Inequalities. 2nd Edn, Cambridge University Press. [9] Hill, D. J. (1991) `A Generalization of the Small-gain Theorem for Nonlinear Feedback Systems', Automatica, 27, 1047±1050. [10] Hill, D. J. and Moylan, P. (1976) `The Stability of Nonlinear Dissipative Systems', IEEE Trans. Automat. Contr., 21, 708±711. [11] Hill, D. J. and Moylan, P. (1977) `Stability Results for Nonlinear Feedback Systems', Automatica, 13, 377±382. [12] Hill, D. J. and Moylan, P. (1980) `Dissipative Dynamical Systems: Basic Input± Output and State Properties', J. of The Franklin Institute, 309, 327±357.
Adaptive Control Systems
157
[13] Ioannou, P. A. and Sun, J. (1996) Robust Adaptive Control. Prentice-Hall, Upper Saddle River, NJ. [14] Jiang, Z. P. and Hill, D. J. (1997) `Robust Adaptive Nonlinear Regulation with Dynamic Uncertainties', Proc. 36th IEEE Conf. Dec. Control, San Diego, USA. [15] Jiang, Z. P. and Hill, D. J. (1998) `Passivity and Disturbance Attenuation via Output Feedback for Uncertain Nonlinear Systems', IEEE Trans. Automat. Control, $, 992±997. [16] Jiang, Z. P. and Mareels, I. (1997) `A Small Gain Control Method for Nonlinear Cascaded Systems with Dynamic Uncertainties', IEEE Trans. Automat. Contr., 42, 292±308. [17] Jiang, Z. P. and Praly, L. (1996) `A Self-tuning Robust Nonlinear Controller', Proc. 13th IFAC World Congress, Vol. K, 73±78, San Francisco. [18] Jiang, Z. P. and Praly, L. (1998) `Design of Robust Adaptive Controllers for Nonlinear Systems with Dynamic Uncertainties', Automatica, Vol. 34, No. 7, 825±840. [19] Jiang, Z. P., Hill, D. J. and Fradkov, A. L. (1996) `A Passi®cation Approach to Adaptive Nonlinear Stabilization', Systems & Control Letters, 28, 73±84. [20] Jiang, Z. P., Mareels, I. and Wang, Y. (1996) `A Lyapunov Formulation of the Nonlinear Small-gain Theorem for Interconnected ISS Systems', Automatica, 32, 1211±1215. [21] Jiang, Z. P., Teel, A. and Praly, L. (1994) `Small-gain Theorem for ISS Systems and Applications', Mathematics of Control, Signals and Systems, 7, 95±120. [22] Kanellakopoulos, I., KokotovicÂ, P. V. and Marino, R. (1991) `An Extended Direct Scheme for Robust Adaptive Nonlinear Control', Automatica, 27, 247±255. [23] Kanellakopoulos, I., KokotovicÂ, P. V. and Morse, A. S. (1992) `A Toolkit for Nonlinear Feedback Design', Systems & Control Letters, 18, 83±92. [24] Khalil, H. K. (1996) Nonlinear Systems. 2nd Edn, Prentice-Hall, Upper Saddle River, NJ. [25] KokotovicÂ, P. V. and Sussmann, H. J. (1989) `A Positive Real Condition for Global Stabilization of Nonlinear Systems', Systems & Control Letters, 13, 125± 133. [26] KrsticÂ, M., Kanellakopoulos, I., and KokotovicÂ, P. V. (1995) Nonlinear and Adaptive Control Design. New York: John Wiley & Sons. [27] Kurzweil, J. (1956) `On the Inversion of Lyapunov's Second Theorem on Stability of Motion', American Mathematical Society Translations, Series 2, 24, 19±77. [28] Lin, Y. (1996) `Input-to-state Stability with Respect to Noncompact Sets, Proc. 13th IFAC World Congress, Vol. E, 73±78, San Francisco. [29] Lozano, R., Brogliato, B. and Landau, I. (1992) `Passivity and Global Stabilization of Cascaded Nonlinear Systems', IEEE Trans. Autom. Contr., 37, 1386±1388. [30] Mareels, I. and Hill, D. J. (1992) `Monotone Stability of Nonlinear Feedback Systems', J. Math. Systems Estimation Control, 2, 275±291. [31] Marino, R. and Tomei, P. (1993) `Robust Stabilization of Feedback Linearizable Time-varying Uncertain Nonlinear Systems', Automatica, 29, 181±189. [32] Marino, R. and Tomei, P. (1995) Nonlinear Control Design: Geometric, Adaptive and Robust. Prentice-Hall, Europe.
158
Adaptive nonlinear control: passivation and small gain techniques
[33] Moylan, P. and Hill, D. (1978) `Stability Criteria for Large-scale Systems', IEEE Trans. Autom. Control, 23, 143±149. [34] Ortega, R. (1991) `Passivity Properties for the Stabilization of Cascaded Nonlinear Systems', Automatica, 27, 423±424. [35] Polycarpou, M. M. and Ioannou, P. A. (1995) `A Robust Adaptive Nonlinear Control Design. Automatica, 32, 423±427. [36] Praly, L. and Jiang, Z.-P. (1993) `Stabilization by Output Feedback for Systems with ISS Inverse Dynamics', Systems & Control Letters, 21, 19±33. [37] Praly, L., Bastin, G., Pomet, J.-B. and Jiang, Z. P. (1991) `Adaptive Stabilization of Nonlinear Systems, In Foundations of Adaptive Control (P.V. KokotovicÂ, ed.), 347± 433, Springer-Verlag. [38] Rodriguez, A. and Ortega, R. (1990) `Adaptive Stabilization of Nonlinear Systems: the Non-feedback Linearizable Case', in Prep. of the 11th IFAC World Congress, 121±124. [39] Safanov, M. G. (1980) Stability and Robustness of Multivariable Feedback Systems. Cambridge, MA: The MIT Press. [40] Sepulchre, R., JankovicÂ, M. and KokotovicÂ, P. V. (1996) `Constructive Nonlinear Control. Springer-Verlag. [41] Seron, M. M., Hill, D. J. and Fradkov, A. L. (1995) `Adaptive Passi®cation of Nonlinear Systems', Automatica, 31, 1053±1060. [42] Sontag, E. D. (1989) `Smooth Stabilization Implies Coprime Factorization', IEEE Trans. Automat. Contr., 34, 435±443. [43] Sontag, E. D. (1989) `Remarks on Stabilization and Input-to-state Stability', Proc. 28th Conf. Dec. Contr., 1376±1378, Tampa. [44] Sontag, E. D. (1990) `Further Facts about Input-to-state Stabilization', IEEE Trans. Automat. Contr., 35, 473±476. [45] Sontag, E. D. (1995) `On the Input-to-State Stability Property', European Journal of Control, 1, 24±36. [46] Sontag, E. D. and Wang, Y. (1995) `On Characterizations of the Input-to-state Stability Property', Systems & Control Letters, 24, 351±359. [47] Sontag, E. D. and Wang, Y. (1995) `On Characterizations of Set Input-to-state Stability', in Prep. IFAC Nonlinear Control Systems Design Symposium (NOLCOS'95), 226±231, Tahoe City, CA. [48] Taylor, D. G., KokotovicÂ, P. V., Marino, R. and Kanellokopoulos, I. (1989) `Adaptive Regulation of Nonlinear Systems with Unmodeled Dynamics', IEEE Trans. Automat. Contr., 34, 405±412. [49] Teel, A. and Praly, L. (1995) `Tools for Semiglobal Stabilization by Partial-state and Output Feedback', SIAM J. Control Optimiz., 33, 1443±1488. [50] Willems, J. C. (1972) `Dissipative Dynamical Systems, Part I: General Theory; Part II: Linear Systems with Quadratic Supply Rates', Archive for Rational Mechanics and Analysis, 45, 321±393. [51] Yao, B. and Tomizuka, M. (1995) `Robust Adaptive Nonlinear Control with Guaranteed Transient Performance', Proceedings of the American Control Conference, Washington, 2500±2504.
7
Active identi®cation for control of discrete-time uncertain nonlinear systems J. Zhao and I. Kanellakopoulos
Abstract The problem of controlling nonlinear systems with unknown parameters has received a great deal of attention in the continuous-time case. In contrast, its discrete-time counterpart remains largely unexplored, primarily due to the diculties associated with utilizing Lyapunov design techniques in a discretetime framework. Existing results impose restrictive growth conditions on the nonlinearities to yield global stability. In this chaper we propose a novel approach which removes this obstacle and yields global stability and tracking for systems that can be transformed into an output-feedback, strict-feedback, or partial-feedback canonical form. The main novelties of our design are: (i) the temporal and algorithmic separation of the parameter estimation task from the control task, and (ii) the development of an active identi®cation procedure, which uses the control input to actively drive the system state to points in the state space that allow the orthogonalized projection estimator to acquire all the necessary information about the unknown parameters. We prove that our algorithm guarantees complete (for control purposes) identi®cation in a ®nite time interval, whose maximum length we compute. Thus, the traditional structure of concurrent on-line estimation and control is replaced by a two-phase control strategy: ®rst use active identi®cation, and then utilize the acquired parameter information to implement any control strategy as if the parameters were known.
160
7.1
Active identi®cation for control of discrete-time uncertain nonlinear systems
Introduction
In recent years, a great deal of progress has been made in the area of adaptive control of continuous-time nonlinear systems [1], [2]. In contrast, adaptive control of discrete-time nonlinear systems remains a largely unsolved problem. The few existing results [3, 4, 5, 6] can only guarantee global stability under restrictive growth conditions on the nonlinearities, because they use techniques from the literature on adaptive control of linear systems [7, 8]. Indeed, it has recently been shown that any discrete-time adaptive nonlinear controller using a least-squares estimator cannot provide global stability in either the deterministic [9] or the stochastic [10] setting. The only available result which guarantees global stability without imposing any such growth restrictions is found in [11], but it only deals with a scalar nonlinear system which contains a single unknown parameter. The backstepping methodology [1], which provided a crucial ingredient for the development of solutions to many continuous-time adaptive nonlinear problems, has a very simple discrete-time counterpart: one simply `looks ahead' and chooses the control law to force the states to acquire their desired values after a ®nite number of time steps. One can debate the merits of such a deadbeat control strategy [12], especially for nonlinear systems [13], but it seems that in order to guarantee global stability in the presence of arbitrary nonlinearities, any controller will have to have some form of prediction capability. In the presence of unknown parameters, however, it is impossible to calculate these `look-ahead' values of the states. Furthermore, since these calculations involve the unknown parameters as arguments of arbitrary nonlinear functions, no known parameter estimation method is applicable, since all of them require a linear parametrization to guarantee global results. This is the biggest obstacle to providing global solutions for any of the more general discrete-time nonlinear problems. In this chapter we introduce a completely dierent approach to this problem, which allows us to obtain globally stabilizing controllers for several classes of discrete-time nonlinear systems with unknown parameters, without imposing any growth conditions on the nonlinearities. The major assumptions are that the unknown parameters appear linearly in the system equations, and that the system at hand can be transformed, via a global parameter-independent dieomorphism, into one of the canonical forms that have been previously considered in the continuous-time adaptive nonlinear control literature [1]. Another major assumption is that our system is free of noise; this allows us to replace the usual least-squares parameter estimator with an orthogonalized projection scheme, which is known to converge in ®nite time, provided the actual values of the regressor vector form a basis for the regressor subspace. The main diculty with this type of estimator is that in general there is no way to guarantee that this basis will be formed in ®nite time. The ®rst steps towards
Adaptive Control Systems
161
removing this obstacle were taken in preliminary versions of this work [14, 15]. In those papers we developed procedures for selecting the value of the control input during the initial identi®cation period in a way that drives the system state towards points in the state space that generate a basis for this subspace in a speci®ed number of time steps. In this chapter we integrate those procedures with the orthogonalized projection estimator to construct a true active identi®cation scheme, which produces a parameter estimate in a familiar recursive (and thus computationally ecient) manner, and at each time instant uses the current estimate to compute the appropriate control input. As a result, we guarantee that all the parameter information necessary for control purposes will be available after at most 2nr steps for output-feedback systems and n 1r steps for strict-feedback systems, where n is the dimension of the system and r is the dimension of the regressor subspace. If the number of unknown parameters p is equal to r, as it would be in any well-posed identi®cation problem, this implies that at the end of the active identi®cation phase the parameters are completely known. If, on the other hand, p r, then we only identify the projection of the parameter vector that is relevant to the system at hand, and that is all that is necessary to implement any control algorithm. In essence, our active identi®cation scheme guarantees that all the conditions for persistent excitation will be satis®ed in a ®nite time interval: in the noise-free case and for the systems we are considering, all the parameter information that could be acquired by any identi®cation procedure in any amount of time, will in fact be acquired by our scheme in an interval which is made as short as possible, and whose upper bound is computed a priori. The fact that our scheme attempts to minimize the length of this interval is important for transient performance considerations, since this will prevent the state from becoming too large during the identi®cation phase. Once this active identi®cation phase is over, the acquired parameter information can be used to implement any control algorithm as if the parameters were completely known. As an illustration, in this chapter we use a straightforward deadbeat strategy. The fact that discrete-time systems (even nonlinear ones) cannot exhibit the ®nite escape time phenomenon, makes it possible to delay the control action until after the identi®cation phase and still be able to guarantee global stability.
7.2
Problem formulation
The systems we consider in this section comprise all systems that can be transformed via a global dieomorphism to the so-called parametric-output-
162
Active identi®cation for control of discrete-time uncertain nonlinear systems
feedback form:
x1 t 1 x2
t T 1
x1
t .. . xn 1
t 1 xn
t T n 1
x1
t
71
xn
t 1 u
t T n
x1
t y
t x1
t where 2 Rp is the vector of unknown constant parameters and i , i 1 . . . n are known nonlinear functions. The name `parametric-output-feedback form' denotes the fact that the nonlinearities i that are multiplied by unknown parameters depend only on the output y x1 , which is the only measured variable; the states x2 . . . xn are not measured. It is important to note that the functions i are not restricted by any type of growth conditions; in fact, they are not even assumed to be smooth or continuous. The only requirement is that they take on ®nite values whenever their argument x1 is ®nite; this excludes 1 , for example, but it is necessary since we want to nonlinearities like x1 1 obtain global results. This requirement also guarantees that the solutions of (7.1) (with any control law that remains ®nite for ®nite values of the state variables) exist on the in®nite time interval, i.e. there is no ®nite escape time. Furthermore, no restrictions are placed on the values of the unknown constant parameter vector or on the initial conditions. However, the form (7.1) already contains several structural restrictions: the unknown parameters appear linearly, the nonlinearities are not allowed to depend on the unmeasured states, and the system is completely noise free: there is no process noise, no sensor noise, and no actuator noise. Our control objective consists of the global stabilization of (7.1) and the global tracking of a known reference signal yd
t by the output x1
t. For notational simplicity, we will denote it i
x1
t for i 1 . . . n. 7.2.1 A second-order example To illustrate the diculties present in this problem, let us consider the case when the system (7.1) is of second order, i.e. x1
t 1 x2
t T 1t x2
t 1 u
t T 2t
72
y
t x1
t Even if were known, the control u
t would only be able to aect the output
Adaptive Control Systems
163
x1 at time t 2. In other words, given any initial conditions x1
0 and x2
0, we have no way of in¯uencing x1
1 through u
0. The best we can do is to drive x1
2 to zero and keep it there. The control would simply be a deadbeat controller, which utilizes our ability to express future values of x1 as functions of current and past values of x1 and u: x1
t 2 x2
t 1 T 1t1 u
t T 2t 1t1 u
t T 2t 1
x2
t T 1t u
t T 2t 1
u
t 1 T
2t
Thus, the choice of control u
t yd
t 2 T 2t 1t1 yd
t 2 T 2t 1 u
t
1 T
2t
1
1
1t
1t t 1
73
74
would yield x1
t yd
t for all t 3 and would achieve the objective of global stabilization. We emphasize that here we use a deadbeat control law only because it makes the presentation simpler. All the arguments made here are equally applicable to any other discrete-time control strategy, as is the parameter information supplied by our active identi®cation procedure. We hasten to add, however, that, from a strictly technical point of view, deadbeat control is perfectly acceptable in this case, for the following two reasons: (1) The well-known problems of poor inter-sample behaviour resulting from applying deadbeat control to sampled-data systems do not arise here, since we are dealing with a purely discrete-time problem. (2) Deadbeat control can result to instability when applied to general polynomial nonlinear systems. As an example, consider the system x1
t 1 x2
t x1
tx22
t x2
t 1 u
t
75
y
t x1
t If we implement a deadbeat control strategy to track the reference signal yd
t 2 t , one of the two possible closed-form solutions yields
76 x2
t 2t 1 22t which is clearly unbounded. The computational procedures presented in [13] provide ways of avoiding such problems. However, in the case of systems of the form (7.1) and of all the other forms we deal with in this
164
Active identi®cation for control of discrete-time uncertain nonlinear systems
chapter, such issues do not even arise, owing to the special structure of our systems which guarantees that boundedness of x1 xi automatically ensures boundedness of xi1 , since xi1
t xi
t 1 T i
x1
t. Of course, when is unknown, the controller (7.4) cannot be implemented. Furthermore, it is clear that any attempt to replace the unknown with an estimate ^ would be sti¯ed by the fact that appears inside the nonlinear function 1 . Available estimation methods cannot provide global results for such a nonlinearly parametrized problem, except for the case where 1 is restricted by linear growth conditions. 7.2.2 Avoiding the nonlinear parametrization Our approach to this problem does not solve the nonlinear parametrization problem; instead, it bypasses it altogether. Returning to the control expression (7.4), we see that its implementation relies on the ability to compute the term T
2t 1t1
77
Since this computation must happen at time t, the argument x1
t 1 is not yet available, so it must be `pre-computed' from the expression x1
t 1 x2
t T 1t u
t
1 T
2t
1
1t
78
Careful examination of the expressions (7.4)±(7.8) reveals that our controller would be implementable if we had the ability to calculate the projection of the unknown parameter vector along known vectors of the form 2
x 1
~ x
79
since then we would be able at time t to compute the terms T
2t
1t
710
T
2t 1t1
711
1
and from them the control (7.4). Hence, our main task is to compute the projection of along vectors of the form (7.9). To achieve this, we proceed as follows: First, we de®ne the subspace spanned by all vectors of
the form (7.9): D
S0 Rf1
x 2
~ x
8 x 2 R 8 x~ 2 Rg
712
Note that the known nonlinear functions 1 and 2 need to be evaluated independently over all possible values of their arguments. This is necessary because we are not imposing any smoothness or continuity assumptions on
Adaptive Control Systems
165
these functions. However, for any reasonable nonlinearities, determining this subspace will be a fairly straightforward task which of course can be performed o-line. The dimension of S0 , denoted by r0 , will always be less than or equal to the number of unknown parameters p: r0 p. In fact, in any reasonably posed problem we will have r0 p, since r0 p means that we are considering more parameters than are actually entering the system equations; in that case, complete parameter identi®cation cannot be achieved with any method or input, since the regressor vector cannot acquire the values necessary to identify some of the parameters. Hence, if r0 p, then the number of unknown parameters can be reduced to r0 without any loss of information or generality.
Clearly, in order to be able to implement the control (7.4), all we need to know about is its projection on the subspace S0 . But how do we acquire this projection? From (7.3) we see that at time t, using the measurements x1
t x1
t 1 x1
t 2 and the known value of the control u
t 2, we can compute the following projection: T 2
x1
t
2 1
x1
t
1 x1
t
u
t
2
713
Hence, if the values of x1 are such that the corresponding values of the vector 2
x1
t 2 1
x1
t 1 eventually form a basis for the subspace S0 , we will obtain all the necessary information about . But how do we guarantee that this identi®cation phase will be of ®nite duration? Instead of allowing the system state to drift on its own, we use the control input u to drive the output x1 to values which result in linearly independent vectors 2t 2 1t 1 and form a basis for S0 in at most 2nr0 steps (where n is the dimension of the system state and r0 the dimension of the nonlinearity subspace). But how can we determine the values of u that will result in such basis vectors in the presence of unknown parameters? This seemingly hopeless dilemma can be resolved by the following observation, which will be clari®ed further later on: The expression (7.4) is not computable if and only if at least one of the vectors 2t 1 1t and 2t 1t1 is independent of the past values 2 j 1 1 j j t 1. Thus, inability to compute (7.4) from already measured projections is equivalent to the knowledge that new independent directions are being generated by the system. In other words, whenever our identi®cation process gets `stuck', that is, the system does not generate new directions over the next few steps, then the projection information we have already acquired is enough for us to compute a value of control which will get the system `unstuck' and will generate a new direction after at most 2n (in this case 4) steps: this is the time it takes to change the arguments of both 1 and 2 and measure the resulting projection.
166
Active identi®cation for control of discrete-time uncertain nonlinear systems
All the projection information of is automatically incorporated into the parameter estimate ^ produced by an orthogonalized projection algorithm. This means that after the active identi®cation phase is complete, all the terms appearing in (7.10) and (7.11) can be ^ This allows us to proceed computed simply by replacing by its estimate . with the implementation of the controller (7.4) or any other control strategy as if the parameters were known. Clearly, this two-stage process depends critically on the fact that, contrary to their continuous-time counterparts, discrete-time nonlinear systems cannot exhibit ®nite escape times, as long as their nonlinearities take on ®nite values whenever their arguments are ®nite. This property allows us to postpone closing the loop with a controller until after the ®nite-duration identi®cation phase has been completed.
7.3
Active identi®cation
Let us now elaborate further on the above outlined approach by presenting in detail its two most challenging ingredients, namely the pre-computation scheme and the input selection for active identi®cation. To do this, we return to the general output-feedback form (7.1) and rewrite it in the following scalar form: x1 t n x2
t n
1 T 1
x1
t n
1
T
x3
t n 2 2
x1
t n 2 T 1
x1
t n .. . n X u
t T k
x1
t n k
1
714
k1
Hence, the following choice of a deadbeat control law: u
t yd
t n
n X
T ktn
k
715
k1
will globally stabilize the system (7.1) and yield x1
t yd
t t n. Clearly, the implementation of the control law (7.15) requires n us to calculate (at time t) the projection of the unknown along the vector k1 ktn k . This n at time t. Rewriting means that we need to compute the value of ktn k k1 n k1 ktn k as n n k
x1
t n k
716 ktn k k1
k1
we can therefore infer that it is necessary for us to be able to calculate the value
Adaptive Control Systems
167
of the states x1 t 1 x1
t n 1 at time t. To see how to calculate these states, let us return to equation (7.14) and express x1
t 1 x1
t n 1 as x1
t i u
t
n i T
n X
k
x1
t i
k
i 1 . . . n
1
717
k1
Clearly, equation (7.17) shows that the value of x1
t 1 depends on the values n T of both k1 kt1 k and u
t n 1. Since the values of u
t n 1 and the vector nk1 kt1 k are known at time t, the key to successfully calculating (at time t) the value of x1
t 1 depends on whether we are able to compute the projection of the unknown along the vector nk1 kt1 k at time t. Next, let us examine what we need to calculate the value of x1
t 2 at time t. From (7.17), the value of x1
t 2 is equal to the sum of u
t n 2 and T nk1 kt2 k . Clearly, if we are able to calculate the values of both u
t n 2 and T nk1 kt2 k at time t, then the value of x2
t 2 can be acquired (at time t). The value of u
t n 2 is known at time t, while from the expression n k1
kt2
k
1t1
n
kt2
k
1
x1
t 1
n 1
k1
x1
t 1
k
k1
k2
718
we see that the value of nk1 kt2 k depends on x1
t 1. This means that pre-computing the value of x2
t 2 requires the values of both x1
t 1 and T nk1 kt2 k . In view of the discussion of the previous paragraph, the calculation of x1
t 1 at time t requires us to compute (at time t) the value of T nk1 kt1 k . Thus, in summary, the calculation of x1
t 2 requires us to pre-compute (at time t) the values of
T nk1 kt1 k
719 T nk1 kt2 k
Generalizing the argument of the previous two paragraphs, we can conclude that the pre-computation of the value of x1
t l (1 l n 1) requires knowledge (at time t) of the vector
T
n
k1
kt1
k
.. .
T nk1 ktl
k
720
168
Active identi®cation for control of discrete-time uncertain nonlinear systems
Hence, the knowledge (at time t) of the above vector with l n T n k1 kt1 k ..
. n T k1 ktn 1 k
1
721
enables us to determine the values of the states x1
t 1 . . . x1
t n 1 at time t. However, the implementation of the control law (7.15) requires the value of T nk1 ktn k also. This leads ®nally to the conclusion that, in order to use the control law (7.15), we need to pre-compute (at time t) the vector T n k1 kt1 k ..
722
. n T k1 ktn k
The procedure for acquiring (7.22) at time t (t n) is given in the next two sections, whose contents can be brie¯y summarized as follows: ®rst, the section on pre-computation explains how to pre-compute the vector (7.22) after the identi®cation phase is complete, that is, for any time t such that St0 1 S0 , where St0 1 denotes the subspace that has been identi®ed at time t (with t n): n D ki k i n . . . t
723 St0 1 R k1
while S0 denotes the subspace formed by all possible values of the regressor vector: n D n 0 D
724 i
zi 8
z1 . . . zn 2 R r0 dim S0 S R i1
Then, the section on input selection for identi®cation shows how to guarantee that the identi®cation phase will be completed in ®nite time, that is, how to ensure the existence of a ®nite time tf at which St0f 1 S0 . In addition, we will show that tf 2nr0 . The reason for this seemingly inverted presentation, where we ®rst show what to do with the results of the active identi®cation and then discuss how to obtain these results, is that it makes the procedure easier to understand. 7.3.1 Pre-computation of projections The pre-computation of (7.22) is implemented through an orthogonalized projection estimator. Therefore, we ®rst review brie¯y the standard version of this estimation scheme; for more details, the reader is referred to Section 3.3 of [7].
Adaptive Control Systems
169
Consider the problem of estimating an unknown parameter vector from a simple model of the following form: yt
t
1T
725
where y
t denotes the (scalar) system output at time t, and
t 1 denotes a vector that is a linear or nonlinear function of past measurements Y
t
1 fy
t
1 y
t
2 . . .g
U
t
1 fu
t
1 u
t
2 . . .g
726
The orthogonalized projection algorithm for (7.25) starts with an initial estimate ^0 and the p p identity matrix P 1 , and then updates the estimate ^ and the covariance matrix P for t 1 through the recursive expressions: Pt 2 t 1 T ^
y
t T ^t 1 T t 1 t 1 if t 1 Pt 2 t 1 6 0 P t 2 t 1 ^ t 1
727 t ^ T if t 1 Pt 2 t 1 0 t 1 Pt 2 t 1 T t 1 Pt 2 Pt 2 if T t 1 Pt 2 t 1 6 0 T P t 1 t 2 t 1 Pt 1
728 Pt 2 if T t 1 Pt 2 t 1 0
This algorithm has the following useful properties, which are given here without proof:
(i) Pt 1
t is a linear combination of the vectors
1 . . .
t. (ii) N Pt Rf
1 . . .
tg. In other words, Pt x 0 if and only if x is a linear combination of the vectors
1 . . .
t. (iii) Pt 1
t?Rf
1 . . .
t 1g. ~ (iv)
t?Rf
1 . . .
t 1g, where ~t ^t is the parameter estimation error. It is worth noting that the orthogonalized projection algorithm produces an estimate ^t that renders ^ Jt
t
y
k
T ^
k
12 0
729
k1
and thus minimizes this cost function. This implies that orthogonalized projection is actually an implementable form of the batch least-squares algorithm, which minimizes the same cost function (7.29). However, the batch least-squares algorithm, that is, the least-squares algorithm with in®nite initial covariance (P 11 0), relies on the necessary conditions for optimality,
170
Active identi®cation for control of discrete-time uncertain nonlinear systems
namely t ^ Jt
k 0 ) ^
1
k
1T ^
t
y
k
k
1
730
k1
k1
which cannot be solved to produce a computable parameter estimate before enough linearly independent measurements have been collected to make the t matrix 1
k 1T invertible. In contrast, the orthogonalized k1
k projection algorithm produces an estimate which, at each time t, incorporates all the information about the unknown parameters that has been acquired up to that time. Now we are ready to describe how to apply the orthogonalized projection algorithm (7.27)±(7.28) to our output-feedback system (7.1). At each time t, we can only measure the output x1
t. Utilizing this measurement and (7.14), we can compute the projection of the unknown vector along the known vector n k1 kt k , that is, n xt T k
x1
t k T
731 t 1 k1
D
D
n
where t 1 k1 kt k and xt x1
t u
t n for any t with t n. Since equation (7.31) is in the same form as (7.25), we can use the expressions (7.27)±(7.28) to recursively construct the estimate ^t and the covariance matrix Pt (with t n 1 in order for (7.31) to be valid), starting from initial estimates ^n and Pn 1 I:1 P T ^ ^t 1 T t 2 t 1
xt T t 1 t 1 if t 1 Pt 2 t 1 6 0 P ^ t 2 t 1 t 1
732 t ^ T t 1 if t 1 Pt 2 t 1 0 T Pt 2 Pt 2 t 1 t 1 Pt 2 if T Pt 2 t 1 6 0 t 1 T t 1 Pt 2 t 1 Pt 1
733 if T Pt 2 t 1 Pt 2 t 1 0 !" When the estimation algorithm (7.32)±(7.33) is applied to the system (7.31), the following properties are true for t n: 0 T t Pt 1 t 0 , t 2 St
t 2
St0 1
) ^t1 ^t
734
1
Pt Pt
1
735
1 This notation is used in place of the traditional ^0 and P 1 to emphasize the fact that for the ®rst n time steps we cannot produce any parameter estimates.
Adaptive Control Systems
v 2 St0
1
T T ) ^t v ^tl v T v
l 0 1 2 . . .
171
736
The proof of this lemma is given in the appendix. The properties (7.34)± (7.36) are crucial to our further development, so let us understand what they mean. Properties (7.34) and (7.35) state that whenever the regressor vector t is linearly dependent on the past regressor vectors, then our estimator does not change the value of the parameter estimate and the covariance matrix; this is due to the fact that the new measurement provides no new projection information. Property (7.35) states that the estimate ^t produced at time t is exactly equal to the true parameter vector , when both are projected onto the subspace spanned by the regressor vectors used to generate this estimate, namely the subspace St0 1 Rf0 . . . t 1 g. This property is one of the cornerstones on which we develop our pre-computing methodology in the following sections, because it implies that at time t we know the projection of the true parameter vector along the subspace St0 1 .
In order to better explain the pre-computation part of our algorithm, we postulate that there exists a ®nite time instant tf n at which the regressor vectors that have been measured span the entire subspace generated by the nonlinearities: St0f
1
S0
737
Hence, at time tf the identi®cation procedure is completed, because the orthogonalized projection algorithm has covered all r0 independent directions of the parameter subspace, and hence has identi®ed the true parameter vector . It is important to note (i) that the existence of such a time tf 2nr0 will be guaranteed through appropriate input selection as part of our active identi®cation procedure in the next section, and (ii) that our ability to compute at such a time tf , be it through orthogonalized projection or through batch least squares, is in fact independent of the manner in which the linearly independent regressor vectors were obtained. The de®nition (7.24) tells us that t 2 S0 St0f
1
8 t tf n
738
Hence, (7.36) implies that T ^tf t T t 8 t tf n
739
In particular, with the help of (7.14) we can then pre-compute (at time t with t tf ) the state x1
t 1 through x1
t 1 u
t 1
n T t u
t 1
T
n ^tf t
740
Using the pre-computed x1
t 1, we can also pre-compute (at time t with
172
Active identi®cation for control of discrete-time uncertain nonlinear systems
t tf ) 1t1 . . . nt1 T 1t1 1 u
t 1 n ^tf t . .. .
. .
T ^ n u
t 1 n tf t nt1 and the vector
t1 1t1
n
it2
741
i
i2
n 1 T i1t1 n ^tf t
1 u
t 1
742
i
i1
Since St0f 1 S0 , the pre-computed t1 still belongs to St0f 1 . Hence, we can repeat the argument from (7.40) to (7.42) to pre-compute (at time t tf ) the state x1
t 2 as x1
t 2 u
t 2 n T t1 u
t 2
T n ^tf t1
743
Then, using the pre-computed x1
t 1 and x1
t 2, we can calculate (at time t tf ) the vectors 1t2 . . . nt2 through T 1t2 1 u
t 2 n ^tf t1 . .. .
744
. . T n u
t 2 n ^tf t1 nt2
and also the vector t2 as follows: t2 1t2 2t1
n
it3
i
i3
2 j1
j u
t 3
j
T
n ^tf t2
j
n 2
In general, since the pre-computed vector (1 l n tl
1
2 S0 St0f
i2t1
i
745
i1
1) satis®es
746
1
we can pre-compute (at time t tf ) the state x1
t l as x1
t l u
t l
n T tl
1
u
t l
T n ^tf tl
1
747
Then, using the pre-computed x1
t 1 . . . x1
t l, we can pre-compute
Adaptive Control Systems
173
(still at time t tf ) the vectors: T 1tl 1 u
t l n ^tf tl 1 . .. .
748
. . T n u
t l n ^tf tl 1 ntl and also tl ni1 itl1 i . In summary, for any t with t tf , using the procedure from (7.46)±(7.48), we can pre-compute the vectors itl , with i 1 . . . n and l 1 . . . n 1 and, thus, the vectors t1 . . . tn 1 2 St0f 1 . Combining this with (7.36), we can pre-compute the vector (7.22) as T T t ^tf t .. ..
749 .
. T ^ tn 1 T tn 1
tf
which implies that after time tf we can implement any control algorithm as if the parameter vector were known.
7.3.2 Input selection for identi®cation So far, we have shown how to pre-compute the values of the future states and the vectors associated with these future states, provided that we can ensure the existence of a ®nite time instant tf n at which St0f 1 S0 . Now we show how to guarantee the existence of such a time tf ; this is achieved by using the control input u to drive the output x1 to values that yield linearly independent directions for the vectors i . This input selection takes place whenever necessary during the identi®cation phase, that is, whenever we see that the system will not produce any new directions on its own. The main idea behind our input selection procedure is the following: At time t, we can determine whether any of the regressor vectors t t1 . . . tn 1 will be linearly independent of the vectors we have already measured. If they are not, then we can use our current estimate ^t and the equation (7.15) to select a control input u
t to drive x1
t n to a value that will generate a linearly independent vector tn . In the worst-case scenario, we will have to use u
t u
t 1 . . . u
t n 1 to specify the values x1
t n x1
t n 1 . . . x1
t 2n 1 in order to generate a linearly independent vector t2n 1 .
!" As long as there are still directions in S0 along which the projection of is unknown, it is always possible to choose the input u so that a
174
Active identi®cation for control of discrete-time uncertain nonlinear systems
new direction is generated after at most 2n steps: dim St0
1
0 dim S0 r0 ) dim St2n
1
dim St0
1
1 8t n
750
Proof The proof of this proposition actually constructs the input selection algorithm. Let us ®rst note that (7.23) yields (for t n) Sn0
1
Sn0 St0
1
S0
751
which implies dim Sn0
1
dim Sn0 dim St0
0 The algorithm that guarantees dim St2n follows:
Step 1
1
dim S0 r0
dim St0
1
1
752
1 is implemented as
At time t, measure x1
t and compute 1t and t .
0 Case 1.1 If T t Pt 1 t 6 0, then by (7.34) we have t 62 St 1 , and therefore 0 0 dim St dim St 1 1. No input selection is needed; return to Step 1 and wait for the measurement of x1
t 1. 0 Case 1.2 If T t Pt 1 t 0, then t 2 St
and St0 St0 1 . Go to Step 2.
1
Step 2 Since t 2 St0 1 , use the procedure (7.38)±(7.42) to calculate all of the following quantities, whose values are independent of u
t (since u
t aects only x1
t n): x1
t 1 1t1 . . . nt1 t1
753
0 0 0 Case 2.1 If T t1 Pt 1 t1 6 0, then t1 62 St 1 and dim St1 dim St 1 1. No input selection is needed; return to Step 1 and wait for the measurement of x1
t 2.
Case 2.2
0 If T t1 Pt 1 t1 0 then t1 2 St
1
0 and St1 St0 1 . Go to Step 3.
Step i (3 i n Since ti 1 2 St0 1 , use the procedure (7.38)±(7.42) to calculate all of the following quantities (whose values are also independent of u
t):
754 x1
t i 1 1ti 1 . . . nti 1 ti 1 0 0 Case i.1 If T ti 1 Pt 1 ti 1 6 0, then ti 1 62 St 1 and dim Sti 1 dim St0 1 1. No input selection is needed; return to Step 1 and wait for the measurement of x1
t i.
Case i.2 If T ti 1 Pt 1 ti Step i 1.
1
6 0, then ti
1
2 St0
1
0 and Sti
1
St0 1 . Go to
Adaptive Control Systems
Step n+1
175
At this step, we have pre-computed all of the following quantities: x1 t 1 x1
t n 1 t1 tn 1 1t1 1tn 1
755 nt1 ntn 1
and we know that the pre-computed vectors satisfy t 2 St0 1 tn
1
2 St0
756
1
Case
n 1.1 If there exists a real number a11 such that T a11 Pt 1 a11 6 0, that is, a11 62 St0 1 ,where n D a11 1
a11 itn1 i
757 i2
then choose the control input u
t to be u
t a11
T ^t tn
758
1
This choice yields x1
t n u
t T tn a11
^T t tn
1 1
T tn
1
a11
759
where the last equality follows from (7.56) and (7.36). Therefore, we have 0 tn a11 62 St0 1 and, hence, dim Stn dim St0 1 1. Return to Step 1 and wait for the measurement of x1
t n 1. If there exists no a11 that renders T a11 Pt 1 a11 6 0, that is, if
Case
n 1.2 1
a
n
jtn1
j2
j
1
a
n
j
x1
t n 1
j 2 St0
1
8a 2 R
j2
760
then go to Step n 2. Step n+2 We have pre-computed all the quantities in (7.55), and we also know from (7.60) that any choice of u
t will result in tn 2 St0 1 . Case
n 2.1
If there exist two real numbers a21 a22 such that
176
Active identi®cation for control of discrete-time uncertain nonlinear systems
0 T a21 a22 Pt 1 a21 a22 6 0, that is, a21 a22 62 St 1 , where D
a21 a22 1
a21 2
a22
n
itn2
761
i
i3
then choose the control inputs u
t and u
t 1 as u
t a21
T ^t tn
u
t 1 a22
T ^t tn
762
1
763
In view of (7.14), these choices yield x1
t n u
t T tn ^T t tn
a11
1 1
T tn
1
a21 x1
t n 1 u
t 1 T tn T ^T t tn tn
a22 a22
764
where we have used (7.36) and the fact that tn 1 tn 2 St0 1 . Therefore, we 0 dim St0 1 1. Return to have tn1 a21 a22 62 St0 1 and, hence, dim Stn1 Step 1 and wait for the measurement of x1
t n 2. Case
n 2.2 2 j1
If no such a21 a22 exist, that is, if
n 2 n j
aj j
x1
t n 2 j
aj jtn2 j j3
j 2 St0
1
j3
j1
8
a1 a2 2 R2
765
then go to Step n 3. Step n i
3 i n 1 We have pre-computed all the quantities in (7.55), and we also know from the previous steps that any choice of u
t u
t 1 . . . u
t i 2 will result in tn tn1 . . . tni 2 2 St0 1 . Case
n i.1 If there exist i real numbers ai1 ai2 . . . aii such that 0 T ai1 ...aii Pt 1 ai1 ...aii 6 0, that is, ai1 ...aii 62 St 1 , where ai1 ...aii
i j1
j
aij
n
ji1
jtni
j
766
Adaptive Control Systems
then choose the control inputs ut u
t i u
t j
T ^t tnj
1 aij
177
1 as 2
1ji
767
In view of (7.14), these choices yield x1
t n j
1 T tnj
1 u
t j
^T t tnj 2
aij
2
T
tnj
2
aij 1 j i
768
where we have used (7.36) and the fact that tn 1 . . . tni 2 2 St0 1 . 0 Therefore, we have tni 1 ai1 ...aii 62 St0 1 and, hence, dim Stni 1 0 dim St 1 1. Return to Step 1 and wait for the measurement of x1
t n i. Case
n i.2 i j1
j
aj
If no such ai1 ai2 . . . aii exist, that is, if n
jtni
j
i
j
aj
j
x1
t n i
j 2 St0
1
ji1
j1
ji1
n
769
i
for all
a1 a2 . . . ai 2 R , then go to Step n i 1. Step 2n We have pre-computed all the quantities in (7.55), and we also know from Step 2n 1 that any choice of u
t u
t 1 . . . u
t n 2 will result in tn tn1 . . . t2n 2 2 St0 1 . Since from (7.50) we know that dim St0 1 dim S0 , we conclude that there is at least one vector of the form D
an1 ...ann
n
j
anj
770
j1
that does not belong to St0 1 , that is, such that T an1 ...ann Pt 2 an1 ...ann 6 0. Therefore, choose the control inputs u
t . . . u
t n 1 as u
t j
1 anj
T ^t tnj
2
1jn
771
In view of (7.14), these choices yield x1
t n j
1 u
t j anj
1 T tnj
^T t tnj 2
anj 1 j n
2
T
tnj
2
772
where we have used (7.36) and the fact that tn 1 . . . t2n 2 2 St0 1 . 0 Therefore, we have t2n 1 an1 ...ann 62 St0 1 and, hence, dim St2n 1 0 dim St 1 1. Return to Step 1 and wait for the measurement of x1
t 2n.
178
Active identi®cation for control of discrete-time uncertain nonlinear systems
This completes the input selection procedure as well as the proof. The input selection algorithm is summarized in Figure 7.1. Figure 7.2 provides a graphic description of the relationships between the control inputs, the vectors and , and the output x1 . This graph illustrates the fact that in order to compute the value of x1 t 2 (at time t) we must know the values of 3t 1 , 2t , 1t1 and u
t 1. The pre-computed vectors 3t 1 , 2t and 1t1 further enable us to calculate the vector t2 . On the other hand, if we want to compute x1
t 3 (at time t), then from Figure 7.1 we see that we need (1) the pre-computed t2 2 St0 1 , (2) the pre-computation of 3t , 2t1 , 1t2 , and (3) the value of u
t.
7.4
Finite duration
Using the computational procedure developed in the proof of Proposition 3.1, we can now guarantee that our active identi®cation procedure will have a ®nite duration: # $!" The active identi®cation procedure completely identi®es the projection of the unknown parameter vector along the subspace S0 at time tf 2nr0 2n dim S0
773 Proof Proposition 3.1 shows that each independent direction in S0 takes at most 2n time steps to identify. Since the dimension r0 of the subspace S0 is equal to the number of such independent directions, the active identi®cation procedure will be completed in at most 2nr0 time steps. Once this procedure is completed, we can proceed with the implementation of any control algorithm as if the parameter vector were known.
7.5
Concluding remarks
In this chaper, we have developed a systematic method to achieve global stabilization and tracking for discrete-time output-feedback nonlinear systems with unknown parameters. Our two-phase control strategy bears some resemblance to dual control [16], which not only stabilizes and regulates the system, but also improves the parameter estimates and the future value of the control. First, in the active identi®cation phase, we systematically use the control to drive the states to desired points so that useful projection information about the unknown parameters is obtained. This process of
Adaptive Control Systems ä
Measure x1
k
ä
l9k k9l 1 ä
ä
ä
Compute l
ä
YES ä
Tl Pk 1 l 6 0?
NO ä
l9l 1
ä
YES
l k n?
NO ä
U 9 u
k
ä
ä
U 9 U T u
l
9 a 2 R l n1 s.t. U a ) Tl Pk 1 l 6 0?
NO
n T
ä
ä
l9l 1
Figure 7.1 The input selection algorithm
YES
179
3t
kt
1
1
1
S
t
2
S 2
ä
ä ä
1
x1
t 4 ä
x1
t 3
ä
u
t
2t
t4
ä
ä
x1
t 2
1
t
0 1
t
S 2
t3
ä
t2
1t
0 1
Active identi®cation for control of discrete-time uncertain nonlinear systems 0 1
180
u
t 1
u
t
1t
1t1
1t2
1t3
2t
2t2
2t2
2t3
3t
3t1
3t2
3t3
kt
kt1
kt2
kt3
Figure 7.2 A graphic representation of the pre-computation procedure
active identi®cation is ®nite. Once all the necessary projection information is obtained, we are able to systematically pre-compute future states and the associated projections. Then, in the subsequent control phase, we use this prediction capability to treat the system as completely known; this means that one can apply any control algorithm (the simplest being `deadbeat' control) that globally stabilizes the system and tracks any given bounded reference signal when the parameters are known. The input selection procedure that we proposed here guarantees that the active identi®cation interval will be of ®nite duration. However, it does not provide any guarantees on the transient behaviour of the states during this phase. Clearly, one may be able to exploit the freedom of choice of ut in order to make this phase shorter and smoother. This issue is a topic of current research.
Adaptive Control Systems
181
Appendix Proof of Lemma 3.1 First, let us prove (7.34). (() Since N Pt 1 Rfn 1 t 1 g St0 1 , t 2 St0 Pt 1 t 0. Thus, we have t 2 St0
1
1
implies that
T ) T t Pt 1 t t 0 0
A1
()) Assume that T t Pt 1 t 0. Then, the fact that Pt 1 t is a linear combination of the vectors n 1 . . . t implies that Pt 1 t 2 St0 , that is, there exist constants cn 1 . . . ct such that Pt 1 t
t
A2
ci i
in 1
Hence, using the de®nition (3.23), we can infer Pt 1 t
ct t
t 1
ci i 2 St0
A3
1
in 1
Since t 2 St0
1
St0 , we can decompose the vector t into t v v?
A4
0 where v denotes the component of t which belongs to St0 1 and v? 2 St is the 0 component of t which is orthogonal to St 1 . Then (A.3) can be reorganized as
Pt 1 t
ct v? ct v
t 1
ci i 2 St0
1
A5
in 1
We know that Pt 1 t ?Rfn 1 . . . t 1 g St0 1 . Hence, we conclude that 0 Combining this with the fact that Pt 1 t ct v? ?St 1 . ? Pt 1 t ct n2 v 2 St0 1 (from (A.5)) yields Pt 1 t that is
ct v? 0
Pt 1 t ct v?
A7
Multiplying both sides of equation (A.7) by T t , we obtain T ? T t Pt 1 t ct t v
A8
In view of the decomposition (A.4), we can simplify (A.8) and obtain T ? T t Pt 1 t ct t v T ? ct
v v? v 2 ct kv? k
A9
182
Active identi®cation for control of discrete-time uncertain nonlinear systems
On the other hand, from the assumption we have T t Pt 1 t 0. Thus, equation (A.9) implies 2 ct kv? k 0
A10
from which we can conclude that ct
? n2 v
0
A11
Substituting (A.11) into (A.7) gives Pt 1 t 0, that is t 2 N
Pt 1 St0
A12
1
So, we have shown that 0 T t Pt 1 t 0 ) t 2 St
A13
1
To prove (7.35), we ®rst use (7.34): t 2 St0
1
) T t Pt 1 t 0
A14
^ ^ But, when T t Pt 1 t 0 the update laws (7.32)±(7.33) yield t1 t and Pt Pt 1 . Finally, the proof of (7.36) is as follows: We know that T ~ t ?Rfn 1 . . . t 1 g. Thus, v 2 St0 1 implies that ~t v 0, which can be T rewritten as ^ v T v. On the other hand, v 2 S 0 implies that v 2 S 0 , t 1 T
t
T
tl 1
0 T ~ ^ since St0 1 Stl 1 . Hence, we also conclude that tl v 0, that is, tl v v. Combining these, we obtain T ^T ^T t v v tl v
8 l 0 1 2 . . .
A15
References [1] KrsticÂ, M., Kanellakopoulos, I. and KokotovicÂ, P. V. (1995). Nonlinear and Adaptive Control Design, Wiley-Interscience, NY. [2] Marino, R. and Tomei, P. (1995). Nonlinear Control Design: Geometric, Adaptive and Robust, Prentice-Hall, London. [3] Chen, F.-C. and Khalil, H. K. (1995). `Adaptive Control of a Class of Nonlinear Discrete-time Systems Using Neural Networks', IEEE Transactions on Automatic Control, Vol. 40, 791±801. [4] Song, Y. and Grizzle, J. W. (1993). `Adaptive Output-feedback Control of a Class of Discrete-time Nonlinear Systems', Proceedings of the 1993 American Control Conference, San Francisco, CA, 1359±1364. [5] Yeh, P.-C. and KokotovicÂ, P. V. (1995). `Adaptive Control of a Class of Nonlinear Discrete-time Systems', International Journal of Control, Vol. 62, 302±324.
Adaptive Control Systems
183
[6] Yeh, P.-C. and KokotovicÂ, P. V. (1995). `Adaptive Output-feedback Design for a Class of Nonlinear discrete-time Systems', IEEE Transactions on Automatic Control, vol. 40, 1663±1668. [7] Goodwin, G. C. and Sin, K. S. (1984). Adaptive Filtering, Prediction and Control, Prentice-Hall, Englewood Clis, NJ. [8] Chen, H. F. and Guo, L. (1991). Identi®cation and Stochastic Adaptive Control, BirkhaÈuser, Boston, MA. [9] Guo, L. and Wei, C. (1996). `Global Stability/Instability of LS-based Discrete-time Adaptive Nonlinear Control', Preprints of the 13th IFAC World Congress, San Francisco, CA, July, Vol. K, 277±282. [10] Guo, L. (1997). `On Critical Stability of Discrete-time Adaptive Nonlinear Control', IEEE Transactions on Automatic Control, Vol. 42, 1488±1499. [11] Kanellakopoulos, I. (1994). `A Discrete-time Adaptive Nonlinear System', IEEE Transactions on Automatic Control, Vol. 39, 2362±2364. [12] AÊstroÈm, K. J. and Wittenmark, B. (1984). Computer Controlled Systems, PrenticeHall, Englewood Clis, NJ. [13] NesÏ icÂ, D. and Mareels, I. M. Y. (1998). `Dead Beat Controllability of Polynomial Systems: Symbolic Computation Approaches', IEEE Transactions on Automatic Control, Vol. 43, 162±175. [14] Zhao, J. and Kanellakopoulos, I. (1997). `Adaptive Control of Discrete-time Strictfeedback Nonlinear Systems', Proceedings of the 1997 American Control Conference, Albuquerque, NM, June, 828±832. [15] Zhao, J. and Kanellakopoulos, I. (1997). `Adaptive Control of Discrete-time Output-feedback Nonlinear Systems', Proceedings of the 36th Conference on Decision and Control, San Diego, CA, December, 4326±4331. [16] Feldbaum, A. A. (1965). Optimal Control Systems, Academic Press, NY.
8
Optimal adaptive tracking for nonlinear systems M. Krstic and Z.-H. Li
Abstract We pose and solve an `inverse optimal' adaptive tracking problem for nonlinear systems with unknown parameters. A controller is said to be inverse optimal when it minimizes a meaningful cost functional that incorporates integral penalty on the tracking error state and the control, as well as a terminal penalty on the parameter estimation error. The basis of our method is an adaptive tracking control Lyapunov function (atclf) whose existence guarantees the solvability of the inverse optimal problem. The controllers designed in this chapter are not of certainty equivalence type. Even in the linear case they would not be a result of solving a Riccati equation for a given value of the parameter estimate. Our abandoning of the CE approach is motivated by the fact that, in general, this approach does not lead to optimality of the controller with respect to the overall plant-estimator system, even though both the estimator and the controller may be optimal as separate entities. Our controllers, instead, compensate for the eect of parameter adaptation transients in order to achieve optimality of the overall system. We combine inverse optimality with backstepping to design a new class of adaptive controllers for strict-feedback systems. These controllers solve a problem left open in the previous adaptive backstepping designs ± getting transient performance bounds that include an estimate of control eort, which is the ®rst such result in the adaptive control literature.
8.1
Introduction
Because of the burden that the Hamilton±Jacobi±Bellman (HJB) pde's impose
Adaptive
185
on the problem of optimal control of nonlinear systems, the eorts made over the last few years in control of nonlinear systems with uncertainties (adaptive and robust, see, e.g., Krstic et al., 1995; Marino and Tomei, 1995; and the references therein) have been focused on achieving stability rather than optimality. Recently, Freeman and Kokotovic (1996a, b) revived the interest in the optimal control problem by showing that the solvability of the (robust) stabilization problem implies the solvability of the (robust) inverse optimal control problem. Further extensive results on inverse optimal nonlinear stabilization were presented by Sepulchre et al. (1997). The dierence between the direct and the inverse optimal control problems is that the former seeks a controller that minimizes a given cost, while the latter is concerned with ®nding a controller that minimizes some `meaningful' cost. In the inverse optimal approach, a controller is designed by using a control Lyapunov function (clf) obtained from solving the stabilization problem. The clf employed in the inverse optimal design is, in fact, a solution to the HJB pde with a meaningful cost. In this chapter we formulate and solve the inverse optimal adaptive tracking problem for nonlinear systems. We focus on the tracking rather than the (setpoint) regulation problem because, even when a bound on the parametric uncertainty is known, tracking cannot (in general) be achieved using robust techniques ± adaptation is necessary to achieve tracking. The cost functional in our inverse optimal problem includes integral penalty on both the tracking error state and control, as well as a penalty on the terminal value of the parameter estimation error. To solve the inverse optimal adaptive tracking problem we expand upon the concept of adaptive control Lyapunov functions (aclf) introduced in our earlier paper (Krstic and KokotovicÂ, 1995) and used to solve the adaptive stabilization problem. Previous eorts to design adaptive `linear-quadratic' controllers (see, e.g., Ioannou and Sun, 1995) were based on the certainty equivalence principle: a parameter estimate computed on the basis of a gradient or least-squares update law is substituted into a control law based on a Riccati equation solved for that value of the parameter estimate. Even though both the estimator and the controller independently possess optimality properties, when combined, they fail to exhibit optimality (and even stability becomes dicult to prove) because the controller `ignores' the time-varying eect of adaptation. In contrast, the Lyapunov-based approach presented in this chapter results in controllers that compensate for the eect of adaptation. A special class of systems for which we constructively solve the inverse optimal adaptive tracking problem in this chapter are the parametric strictfeedback systems, a representative member of a broader class of systems dealt with in Krstic et al. (1995), which includes feedback linearizable systems and, in particular, linear systems. A number of adaptive designs for parametric strictfeedback systems are available, however, none of them is optimal. In this
186
chapter we present a new design which is optimal with respect to a meaningful cost. We also improve upon the existing transient performance results. The transient performance results achieved with the tuning functions design in Krstic et al. (1995), even though the strongest such results in the adaptive control literature, still provide only performance estimates on the tracking error but not on control eort (the control is allowed to be large to achieve good tracking performance). The inverse optimal design in this chapter solves the open problem of incorporating control eort in the performance bounds. The optimal adaptive control problem posed here is not entirely dissimilar from the problem posed in the award-winning paper of Didinsky and Bas°ar (1994) and solved using their cost-to-come method. The dierence is twofold: (a) our approach does not require the inclusion of a noise term in the plant model in order to be able to design a parameter estimator, (b) while Didinsky and Bas° ar (1994) only go as far as to derive a Hamilton±Jacobi±Isaacs equation whose solution would yield an optimal controller, we actually solve our HJB equation and obtain inverse optimal controllers for strict-feedback systems. A nice marriage of the work of Didinsky and Bas° ar (1994) and the backstepping design in Krstic et al. (1995) was brought out in the paper by Pan and Bas°ar (1996) who solved an adaptive disturbance attenuation problem for strict-feedback systems. Their cost, however, does not impose a penalty on control eort. This chapter is organized as follows. In Section 8.2, we pose the adaptive tracking problem (without optimality). The solution to this problem is given in Sections 8.3 and 8.4 which generalize the results of Krstic and Kokotovic (1995). Then in Section 8.5 we pose and solve the inverse optimal problem for general nonlinear systems assuming the existence of an adaptive tracking Lyapunov function (atclf). A constructive method for designing atclf's based on backstepping is presented in Section 8.6, and then used to solve the inverse optimal adaptive tracking problem for strict-feedback systems in Section 8.7. A summary of the transient performance analysis is given in Section 8.8.
8.2
Problem statement: adaptive tracking
We consider the problem of global tracking for systems of the form x_ f x Fx gxu y hx
8:1
where x Rn , u R, the mappings f x, Fx, gx and hx are smooth, and is a constant unknown parameter vector which can take any value in Rp . To make tracking possible in the presence of an unknown parameter, we make the following key assumption.
187
(A1) For a given smooth function yr t, there exist functions t; and r t; such that dt; f t; Ft; gt; r t; dt yr t ht; ;
t 0;
Rp
8:2
Note that this implies that @ h t; 0; @
t 0; Rp
8:3
For this reason, we can replace the objective of tracking the signal where t yr t h t; by the objective of tracking yr t h t; t, is a time function Ð an estimate of customary in adaptive control. which is governed by Consider the signal xr t t; t
@t; @t; @t; 8:4 f xr Fxr gxr r t; @t @ @ and compute its We de®ne the tracking error e x xr x t; derivative: x r
e f x f xr gx gxr r t; @t; Fx Fxr gx u r t; @ Fr f F
8:5
@ g u @
where and
f x f xr gx gxr r t; ft; e;
Fx Fxr e; Ft; Fxr Fr t; u u r t;
8:6
The global (With a slight abuse of notation, we will write gx also as gt; e; .) tracking problem is then transformed into the problem of global stabilization of the error system (8.5). This problem is, in general, not solvable with static feedback. This is obvious in the scalar case n p 1 where, even in the case yr t xr t 0, a control law u x independent of would have the impossible task to satisfy x f x Fx gxx < 0 for all x 0 and all R. Therefore, we seek dynamic feedback controllers to stabilize system (8.5) for all .
188
De®nition 2.1 The adaptive tracking problem for system (8.1) is solvable if smooth on e; (A1) is satis®ed and there exist a function t; n p and a 0; 0, a smooth function t; e; , R R 0 R with t; positive de®nite symmetric p p matrix , such that the dynamic controller e; u t;
8:7
t; e;
8:8
guarantees that the equilibrium e 0; 0 of the system (8.5) is globally stable and et 0 as t for any value of the unknown parameter Rp .
8.3
Adaptive tracking and atclf's
Our approach is to replace the problem of adaptive stabilization of (8.5) by a problem of nonadaptive stabilization of a modi®ed system. This allows us to study adaptive stabilization in the Sontag±Artstein framework of control Lyapunov functions (clf) (Sontag, 1983; Artstein, 1983; Sontag, 1989). De®nition 3.1 A smooth function Va R Rn Rp R , positive de®nite, decrescent, and proper (radially unbounded) in e (uniformly in t) for each , is called an adaptive tracking control Lyapunov function (atclf) for (8.1) (or alternatively, an adaptive control Lyapunov function (aclf) for (8.5)), if (A1) is satis®ed and there exists a positive de®nite symmetric matrix Rpp such that for each Rp , Va t; e; is a clf for the modi®ed nonadaptive system @Va @ @Va F g u 8:9 e f F F @ @ @e that is, Va satis®es @Va @Va @Va @ @Va f F F F g u
@Va @Va
@Va @Va 2 @Va 4 > > > f f g > > @e @t @e @e @Va > < @t gt; e; 0 ; @V @e a e; t; g > > @e > > > > @Va > :0; gt; e; 0 @e 8:14 where @Va @ @Va
f f F F F @ @ @e
8:15
With the choice (8.14), inequality (8.13) is satis®ed with the continuous function
190
s 2 4 @Va @Va
@Va t; e; f t; e; gt; e; Wt; e; @t @e @e
8:16
which is positive de®nite in e (uniformly in t) for each , because (8.10) implies that @Va @Va @Va
gt; e; 0 t; e; f t; e; < 0; @e @t @e
e 0 ; t 0 8:17
e; , smooth away from e 0, will be also We note that the control law t; continuous at e 0 if and only if the atclf Va satis®es the following property, called the small control property (Sontag, 1989): for each Rp and for any " > 0 there is a > 0 such that, if e 0 satis®es e , then there is some u with u " such that @Va @Va @Va @ @Va f F F F g u 0; 8:37
satis®es 3 @V1 @ @V1 f F F F g7 @ @ @e @V1 @V1 6 7 6 2 7 W c 6 5 4 @t @ @V @e; 1 r 1 t; e; ; F @ @e 2
8:38
195
e; . With Let us start by introducing for brevity a new error state z t; (8.35) we compute g @V1 @V1 f F @t 1 t; e; ; @e; @Va @ @Va @ @V1 1 z z f F g @t @e @t @e @ @Va @Va @Va @ @ 1 z f F g f F g g @t @e @t @e @e
8:39
On the other hand, in view of (8.35), we have 2
3 @V1 @ @V1 F 7 F @V1 6 @ @ @e 6 7 6 7 4 5 @r @V1 @e; F @ @e # " @ @Va @ @Va @ @Va @ z z F z F F @e @ @ @e @e @ @e @r @V1 F z @ @e @Va @Va @Va @ @Va F F @e @ @e @ @e " # @ @ @r @ @V1 @Va @Va @ @ F F z @ @e @ @e @ @e @ @e @ 8:40 Adding (8.39) and (8.40), with (8.13) and (8.37) we get (8.38).This proves by is an atclf for system (8.33), or, an aclf for (8.34), Theorem 3.1 that V1 t; e; ; and by Corollary 3.1 the adaptive quadratic tracking problem for this system is solvable. The new tuning function is determined by the new atclf V1 and given by
196
@V1 F @V1 @Va @ F F 0 @e @e @e @e; @ F t; e; @e
1 t; e; ;
8:41
in (8.37) is only one out of many possible 1 t; e; ; The control law control laws. Once we have shown that V1 given by (8.35) is an atclf for (8.33), 1 given by or, an aclf for (8.34), we can use, for example, the C0 control law Sontag's formula (8.14). Example 4.1 (Example 3.1 continued)
Let us consider the system:
x 1 x2 'x1 x 2 x3
8:42
x 3 u y x1
We treat the state x3 as an integrator added to the (x1 ; x2 )-subsystem for e; , where Example 3.1, so Lemma 4.1 is applicable. De®ning z3 x3 t; x3 x3 r , by Lemma 4.1, the function V1 t; e; x3 ; 12 z21 z22 z23 is an atclf for the system (8.42). With (8.37) and (8.41) we obtain 1 t; e; ; u @ e2 ' @ @t @e x3 @ @r @ @ 'yr 1 ' ' z2 @ @e2 @e1 @
z2 cz3
1 1 t; e; x3 ;
@ 'z3 @e1
8:43 8:44
@r @'yr yr t actual control is u u r1 where r1 @t @yr @ 2 'yr 2 y t . yr t r @y2r A repeated application of Lemma 4.1 (generalized as in Krstic et al., 1995 (page 138, to the case where u x; ) recovers our earlier result (Krstic et al., 1992). The
Corollary 4.1 [Krstic et al., 1992] The adaptive quadratic tracking problem for the following system is solvable
x_ i xi1 i x1 ; ; xi ;
197
i 1; ; n 1
x n u 'n x1 ; ; xn
8:45
y x1
8.5
Inverse optimal adaptive tracking
While in the previous sections our objective was only to achieve adaptive tracking, in this section our objective is to achieve its optimality in a certain sense. De®nition 5.1 The inverse optimal adaptive tracking problem for system (8.1) is solvable if there exist a positive constant , a positive real-valued function rt; e; , a real-valued function lt; e; positive de®nite in e for each , and a dynamic feedback law (8.7), (8.8) which solves the adaptive quadratic tracking problem and also minimizes the cost functional Z rt; e; u2 dt 2 1 J t lt; e; 8:46 t
0
p
for any R . This de®nition of optimality puts penalty on e and u as well as on the Even though t is not guaranteed to have a limit in the terminal value of . general tracking case (it is guaranteed to have a limit in the case of set-point 2 1 is regulation (KrsticÂ, 1996; Li and KrsticÂ, 1996), the existence of t assured by the assumption that the adaptive quadratic tracking problem is solvable. This can be seen by noting that, since Vt 0 and from (8.22) Vt is nonincreasing, Vt has a limit. Since (8.22) guarantees that Va t 0, it 2 1 has a limit. The absence of an integral penalty on follows from (8.19) that in (8.46) should not be surprising because adaptive feedback systems, in general, do not guarantee parameter convergence to a true value. Theorem 5.1 Suppose there exists an atclf Va t; e; for (8.1) and a control e; that stabilizes the system law u t; @Va @ @Va F e f F F g u 8:47 @ @ @e |{z} f
has the form
e; r 1 t; e; t; where rt; e; > 0 for all t; e; . Then
@Va g @e
8:48
198
(1) The nonadaptive control law t; e; t; e; ; u
2
minimizes the cost functional Z Ja lt; e; rt; e; u2 dt;
8:49
Rp
0
along the solutions of the nonadaptive system (8.47), where 2 @Va @Va
1 @Va 2r lt; e; 2 f g g @t @e @e
8:50
8:51
(2) The inverse optimal adaptive tracking problem is solvable.
In light of (8.13), we have 2 @Va @Va
1 @Va 2r f g g lt; e; 2 @t @e @e 2 1 @Va 2 Wt; e; 2r g @e
Proof (Part 1)
8:52
Since 2, rt; e; > 0, and Wt; e; is positive de®nite, lt; e; is also positive de®nite. Therefore Ja de®ned in (8.50) is a meaningful cost functional, which puts penalty both on e and u. Substituting lt; e; and u r 1 v u
@Va g @e
8:53
into Ja , we get 2 Z @Va @Va
@Va 2 1 @Va f r 2 g rv2 2 v g Ja 2 @t @e @e @e 0 2 @Va 2 r 1 g dt @e Z Z @Va @Va
8:54 f g u dt rv2 dt 2 @t @e 0 0 Z Z dVa rv2 dt 2 0
0
2 Va 0; e0; 0 2 Va t; et; t t
Z
rv2 dt
0
Since the control input ut solves the adaptive quadratic tracking problem,
t et 0, and we have that t Va t; et; t 0. Therefore, the
199
t; e; minimum of (8.54) is reached only if v 0, and hence the control u is an optimal control. (Part 2) Since there exists an atclf Va for (8.1), the adaptive quadratic tracking problem is solvable. Next, we show that the dynamic control law t; e; u
8:55
@Va F t; e; @e
8:56
minimizes the cost functional (8.46). The choice of the update law (8.56) is due to the requirement that (8.55), (8.56) solves the adaptive quadratic tracking and problem (see the proof of Corollary 3.1). Substituting lt; e; v u r 1
@Va g @e
8:57
into J, along the solutions of (8.5) and (8.56) we get @Va @Va @ @Va 2 F f F Fr t @t @e @e @ 0 2 @Va 1 @Va @Va F 2 r 1 g 2 @e @e @ 2 @Va 2 1 @Va 2 g r g rv 2 v dt @e @e Z @Va @Va @ @Va 2 f F Fr F g u 1 2 t @t @e @e @ 0 # Z @Va 1 @Va F rv2 dt dt @e @ 0 Z Z 1 1 2 d Va 1 2 rv2 dt t 2 0 0 Z 2 1 2 Va t; et; t 0 rv2 dt 2 Va 0; e0; 0
2 1 J
Z
2
t
0
8:58
Again, since ut solves the adaptive quadratic tracking problem,
t et 0, and we have that t Va t; et; t 0. Therefore, the t; e; minimum of (8.58) is reached only if v 0, thus the control u minimizes the cost functional (8.46).
200
Remark 5.1 Even though not explicit in the proof of the above theorem, the atclf Va t; e; solves the following family of HJB equations parametrized in 2: # @Va @Va @Va @ @Va f F F F @ @t @e @ @e |{z} 8:59 2 @Va lt; e; 0 g 2rt; e; @e 2 The under-braced terms represent the `non-certainty equivalence' part of this HJB equation. Their role is to take into account the time-varying eect of parameter adaptation and make the control law optimal in the presence of an update law. Remark 5.2 The freedom in selecting the parameter 2 in the control law (8.49) means that the inverse optimal adaptive controller has an in®nite gain margin. Consider the scalar linear system
Example 5.1
x u x
8:60
yx
For simplicity, we focus on the regulation case, yr t 0. Since the system is scalar, Va 12 x2 , Lg Va x, Lf Va x2 . We choose the control law based on Sontag's formula p us x 2 1 2r 1 x 8:61 where
r The control
us 2
2 > 0; 2 1
is stabilizing for the system (8.60) because V a us 12 2 1 x2 2
8:62
8:63
By Theorem 5.1, the control us is optimal with respect to the cost functional Z Z x2 u2 dt Ja lx; ru2 dt 2 8:64 2 1 0 0
with a value function Ja x 2x2 . Meanwhile, the dynamic control
q us x 2 1 x2 is optimal with respect to the cost functional Z x2 u2 2 p dt J 2 2 2 1 0
201
8:65 8:66
8:67
x2 2 . with a value function J x; We point out that exists both due to the scalar (in parameter ) nature of the problem and because it is a problem of regulation (KrsticÂ, 1996). Note that, even though the penalty coecient on x and u in (8.67) varies with t, the penalty coecient is always positive and ®nite. Remark 5.3 The control law (8.61) is, in fact, a linear-quadratic-regulator (LQR) for the system (8.60) when the parameter is known. The control law can be also written as us px where p is the solution of the Riccati equation p2 2p 1 0 8:68 It is of interest to compare the approach in this chapter with `adaptive LQR schemes' for linear systems in Ioannou and Sun (1995, Section 7.4.4). . Even though both methodologies result in the same control law (8.61) for the scalar linear system in Example 5.1, they employ dierent update laws. The gradient update law in Ioannou and Sun (1995, Section 7.4.4) is optimal with respect to an (instantaneous) cost on an estimation error; however, when its are substituted into the control law (8.65), this control law is estimates t not (guaranteed to be) optimal for the overall system. (Even its proof of stability is a non-trivial matter!) In contrast, the update law (8.66) guarantees optimality of the control law (8.65) for the overall system with respect to the meaningful cost (8.67). . The true dierence between the approach here and the adaptive LQR scheme in Ioannou and Sun (1995, Section 7.4.4) arises for systems of higher order. Then the under-braced non-certainty equivalence terms in (8.59) start to play a signi®cant role. The CE approach in Ioannou and Sun (1995) would be to set 0 in the HJB (Riccati ± for linear systems) equation (8.59) and combine the resulting control law with a gradient or least-squares update law. The optimality of the nonadaptive controller would be lost in the In contrast, a solution to presence of adaptation due to the time-varying t. Example (5.1) with > 0 would lead to optimality with respect to (8.46).
Corollary 5.1 If there exists an atclf Va t; e; for (8.1) then the inverse optimal adaptive tracking problem is solvable.
202
e; where t; e; is Proof Consider the Sontag-type control law us t; e; is an asymptotic stabilizing de®ned by (8.14).The control law u2s 12 t; controller for system (8.9) because inequality (8.13) is satis®ed with 2 s 3 @V @V @Va @Va 2 @Va 4 5 a a
f f g 8:69 Wt; e; 12 4 @t @e @t @e @e
@Va @Va
f 0 given by 2 t; 8 @e @Va 2 > > > g 2 > > @Va @e > > s; g 0 > < 2 4 @e @V @V @V @V @V a a
a a
a rt; e; f f g > > @t @e @t @e @e > > > > > @Va > : any positive real number, g0 @e 8:70 by Theorem 5.1, the inverse optimal adaptive tracking problem is solvable. The optimal control is the formula (8.14) itself. Corollary 5.2 The inverse optimal adaptive tracking problem for the following system is solvable x i xi1 'i x1 ; ; xi ;
i 1; ; n 1
x n u 'n x1 ; ; xn
8:71
y x1 Proof By Corollary 4.1 and Corollary 3.1, there exists an atclf for (8.71). It then follows from Corollary 5.1 that the inverse optimal adaptive tracking problem for system (8.71) is solvable.
8.6
Inverse optimality via backstepping
With Theorem 5.1, the problem of inverse optimal adaptive tracking is reduced to the problem of ®nding an atclf. However, the control law (8.14) based on Sontag's formula is not guaranteed to be smooth at the origin. In this section
203
we develop controllers based on backstepping which are smooth everywhere, and, hence, can be employed in a recursive design. Lemma 6.1
If the adaptive quadratic tracking problem for the system x_ f x Fx gxu y hx
8:72
e; and (8.13) is satis®ed with is solvable with a smooth control law t; Wt; e; e t; e; e, where t; e; is positive de®nite and symmetric for all t; e; ; then the inverse optimal adaptive tracking problem for the augmented system x f x Fx gx u
8:73
y hx is also solvable with a smooth control law. Proof Since the adaptive quadratic tracking problem for the system (8.72) is Va t; e; solvable, by Lemma 4.1 and Corollary 3.1, V1 t; e; ; 2 1 e; is an atclf for the augmented system (8.73), i.e. a clf for the 2 t; modi®ed nonadaptive error system (8.36). Adding (8.39) and (8.40), with (8.13), we get 2 3 @V1 @ @V1 F g7 f F F @ @ @e @V1 @V1 6 6 7 V 1 6 7 4 5 @t @r @V1 @e; u F @ @e @Va @ @ z f F g g W z u @t @e @e @ @ @ @ @r @ @Va @r @ @ F z F @ @e @ @ @e @ @e @ @e @ @Va @Va @ @ F @ @e @ @e 8:74
e; . To render V 1 negative de®nite, one choice is (8.37) where z t; which cancels all the nonlinear terms inside the bracket in (8.74). However, the cancellation controller (8.37) is not (guaranteed to be) optimal. Therefore, we have to use other techniques in the design of our control law. One such technique we will use here is `nonlinear damping' (Krstic et al., 1995).
204
@Va @Va @ , f , F, , are smooth and vanish for e 0, then we can Since , @t @e @ write @Va @ @ f F g g @t @e @e @ @ @r @ @Va @Va @Va @ @ F F @ @e @ @e @ @e @ @e @
8:75
1 t; e; t; e; 1=2 e
where 1 t; e; is a vector-valued smooth function and t; e; 1=2 is invertible for all t; e; . In addition, let us denote @ @ @ @r @ @ g F 2 t; e; 8:76 @e @ @e @ @e @ Then (8.74) is re-written as V 1 1=2 e2 z u z1 1=2 e 2 z2 The choice
renders
2 2 c 1 2 z; 1 t; e; ; u 2 2c
c>0
c V 1 12 1=2 e2 z2 2
R 1 t; e; ;
1 2 22 c 2c 2
"
> 0;
8:78 8:79
1 de®ned in (8.78) is of the form Since the control law u @V1 0 1 1 t; e; ; R t; e; ; 1 @e; where
8:77
; t; e; ;
8:80
8:81
by Theorem 5.1, the dynamic feedback control (adaptive control) u 2, with , 1 t; e; ; @V1 Ft; e; ; 8:82 1 t; e; ; @e is optimal for the closed loop tracking error system (8.34) and (8.32). Example 6.1 (Example 4.1 revisited) For the system (8.42), we designed a controller (8.43) which is not optimal due to its cancellation property. With Lemma 6.1, we can design an optimal control as follows. First we note that given by (8.28) in Example 3.1 is of the form
@' @ e; 1 ' ' c1 t; z1 c1 @e1 @t @' c2 c1 1 ' ' z2 @e1
205
8:83
at; e; z1 bt; e; z2
'x1 'xr1 'e1 xr1 'xr1 e1 e1 and @@t' z1 @ because ' @t . Instead of (8.43) we choose the `nonlinear damping' control suggested by Lemma 6.1: 1 t; e; ; u # 2 @ 1 @ @a @ @r @ c1 a ' ' c3 2c1 @e1 @t @e2 @ @e2 r @ @b @ 1 @ @ 1 ' ' b 2c2 @e1 @e1 @t @e2 2 2 $ @ @r @ @' 1 @ z3 ' ' c1 @ @e2 r 2c3 @e2 @e1 @
8:84
The tuning function 1 is the same as in (8.44). The control law 2, with 1 is optimal. , 1 t; e; ; u
8.7
Design for strict-feedback systems
We now consider the parametric strict-feedback systems (Krstic et al., 1995): x i xi1 'i
xi ;
x n u 'n x
i 1; ; n 1 8:85
y x1 where we use the compact notation x i x1 ; ; xi , and develop a procedure for optimal adaptive tracking of a given signal yr t. An inverse optimal design following from Corollary 5.2 (based on Sontag's formula) would be nonsmooth at e 0. In this section we develop a design which is smooth everywhere. This design is also dierent from the nonoptimal design in Krstic et al. (1992, 1995). For the class of systems (8.85), (A1) is satis®ed for any function yr t, and there exist functions 1 t, 2 t; , ,n t; , and r t; such that
206
_ i i1 'i
i ;
i 1; ; n 1
n r t; 'n
8:86
yr t 1 t which is governed by Consider the signal xr t t; x ri xr;i1 'ri
xri
@i ; @
'rn xr x rn r t;
i 1; ; n 1
@n @
8:87
yr xr1 The tracking error e x xr is governed by the system 'ri t; i t; e i ; ei ei1 '
@i ; @
i 1; ; n 1
'rn t; @n n t; e; en u ' @
8:88
and ' i 'i
xi 'ri
xri , i ; ; n. For an atclf Va , the where u u r t; modi®ed nonadaptive error system is @Va @i @Va F ; i 1; ; n 1 ei ei1 @ @e @ @Va @n @Va n t; e; 'n F en u ' @ @e @ ' i
' i
8:89
where F '1 ; ; 'n T . First, we search for an atclf for the system (8.85). Repeated application of Lemma 4.1 gives an atclf Va 12
n %
z2i
i1
8:90
i 1 t; e i 1 ; zi ei i 's are to be determined. For notational convenience we de®ne z0 0, where 0 0. We then have
n % j 1 @Va @ zj @ @ j1
@Va F @e
n % @Va j1
@ej
'j
207
8:91
n % j1
" n n % % k 1 @ wj zj zj zk 'j @ej j1 kj1
8:92
where wj t; e j ; 'j
j 1 % j 1 @ k1
@ek
8:93
'k
Therefore the modi®ed nonadaptive error system (8.89) becomes ei ei1 ' i en u
' n
n % j 1 @
@
j1
n % j 1 @ j1
@
'i zj
'n zj
n % @i j1
@
n % @n j1
@
wj zj ;
i 1; ; n 1 8:94
wj zj
1 , , n 1 are yet to be determined to make Va de®ned in (8.90) The functions a clf for system (8.94). To design these functions, we apply the backstepping technique as in Krstic et al. (1995). We perform cancellations at all the steps before step n. At the ®nal step n, we depart from Krstic et al. (1995) and choose the actual control u in a form which, according to Lemma 6.1, is inverse optimal. Step
i 1; ; n 1: i t; e i ; zi 1 ci zi
i 1 % ki ik zk ii zi ;
ci > 0
k1
i 1 % i 1 @
8:95
k '
8:96
" i 1 i 1 @i % i 1 @j @ @ ik wk @ @ @ej @ j2
8:97
i t; e i ; ' i w
Step n:
i 1 i 1 % i 1 @ @ ek1 w i @t @e k k1
k1
@ek
With the help of Lemma A.1 in the appendix, the derivative of Va is
208
V_ a
n 1 %
ck z2k
k1
"
zn zn 1
n 1 % k1
kn nk zk nn zn u 8:98
# n 1 n 1 X n 1 @ @ n ek1 w @t @ek k1
We are now at the position to choose the actual control u. We may choose u such that all the terms inside the bracket are cancelled and the bracketed term multiplying zn is equal to cn z2n as in Krstic et al. (1995), but the controller designed in that way is not guaranteed to be inverse optimal. To design a controller which is inverse optimal, according to Theorem 5.1, we should choose a control law that is of the form n t; e; r 1 t; e; u
@Va g @e
8:99
where rt; e; > 0, t; e; . In light of (8.94) and (8.90), (8.99) simpli®es to n t; e; r 1 t; e; zn u
8:100
n with zn as a factor. i.e. we must choose k , k 1; ; n 1, and the expression in the second Since ek1 zk1 line in (8.98) vanishes at e 0, it is easy to see that it also vanishes for z 0. Therefore, there exist smooth functions k ; k 1; ; n, such that
n 1 n X n 1 X n 1 @ @ k zk ek1 w n @t @ek k1 k1
8:101
Thus (8.98) becomes V a
n 1 X k1
ck z2k zn u
n X
zn k zk
k1
8:102
where k kn nk k ;
k 1; ; n 2 8:103
n 1 1 n 1;n n;n 1 n 1 n nn n A control law of the form (8.100) with ! 1 n X 2k > 0; rt; e; cn 2ck k1 results in
ck > 0;
t; e;
8:104
V_ a 12
n % k1
ck z2k
n % ck k1
2
k zk zn ck
2
209
8:105
By Theorem 5.1, the inverse optimal adaptive tracking problem is solved through the dynamic feedback control (adaptive control) law 2 n t; e; n t; e; u n % @Va F wj zj @e j1
8.8
8:106
Transient performance
In this brief section, we give an 2 bound on the error state z and control u for the inverse optimal adaptive controller designed in Section 8.7. According to Theorem 5.1, the control law (8.106) is optimal with respect to the cost functional 2 1 J 2 t t 2 2
Z
0
n 6 6X
6 ck z2k 6 4 k1
with a value function
2 k ck zk zn ck k1
n X
3
7 8:107 7 u2 !7 dt 7 n X 2k 5 2 cn 2ck k1
2 1 2z2 J 2
8:108
In particular, we have the following 2 performance result. Theorem 8.1 In the adaptive system (8.88), (8.106), the following inequality holds 3 2 Z
0
n 6 6X
6 ck z2k 6 4 k1
7 7 u2 2 1 z02 !7 dt 0 n X 2 7 5 k 2 cn 2ck k1
8:109
This theorem presents the ®rst performance bound in the adaptive control literature that includes an estimate of control eort. The bound (8.109) depends on z0 which is dependent on the design parameters c1 ; ; cn , and . To eliminate this dependency and allow a
210
systematic improvement of the bound on z2 , we employ trajectory initialization as in Krstic et al. (1995, Section 4.3.2) to set z0 0 and obtain: 3 2 Z
0
8.9
n 6 6X
6 ck z2k 6 4 k1
7 7 u2 2 1 !7 dt 0 n X 2 7 5 k 2 cn 2ck k1
8:110
Conclusions
In this chapter we showed that the solvability of the inverse optimal adaptive control problem for a given system is implied by the solvability of the HJB (nonadaptive) equation for a modi®ed system. Our results can be readily extended to the multi-input case and the case where the input vector ®eld depends on the unknown parameter. In constructing an inverse optimal adaptive controller for strict-feedbacksystems we followed the simplest path of using an atclf designed by the tuning functions method in Krstic et al. (1995). Numerous other (possibly better) choices are possible, including an inverse optimal adaptive design at each step of backstepping. The relative merit of dierent approaches is yet to be established.
Appendix
A technical lemma
Lemma A.1 The time derivative of Va in (8.90) along the solutions of system (8.94) with (8.95)±(8.97) is given by V a
n 1 X k1
ck z2k
zn zn 1
n 1 X kn nk zk nn zn u: k1
n 1 n 1 X n 1 @ @ ek1 w n @t @e k k1
Proof
First we prove that the closed loop system after i steps is
A:1
211
3 c1 1 12 13 1;i 1 1i 6 c2 1 23 2;i 1 2i 7 72 3 2 3 6 1 12 7 z1 6 z_1 6 7 76 7 6 7 6 13 1 23 76 7 6 76 74 5 4 5 6 7 6 1 i 2;i 1 i 2;i 7 zi 6 zi 6 7 4 1;i 1 2;i 1 1 i 2;i 1 ci 1 1 i 1;i 5 2
1i 2i i 2;i 3 2 1;i1 1;i2 1n 2 3 6 2;i2 2n 7 7 zi1 6 2;i1 76 6 76 7 6 6 74 7 5 7 6 7 6 4 i 1;i1 i 1;i2 i 1;n 5 zn 1 i;i1
i;i2
1 i 1;i
ci
i;n
A:2
and the resulting V i is V i
i X k1
ck z2k zi zi1
n X
ji1
zj
i X
kj zk
k1
A:3
where ik ik ik ik
k 1 @ wi @
! i 1 k 1 @j @i X @ ik wk @ @ej @ j2
A:4 A:5 A:6
1 c1 z1 ' The proof is by induction. Step 1: Substituting 1 into (8.94) @1 0 0 and with i 1, using (8.90) and noting 0, we get @ n X j 1 @ w1 c1 z1 z2 12 z3 1n zn A:7 zj z1 c1 z1 z2 @ j1 and V 1 c1 z21 z1 z2
n X j1
zj
n X j 1 @ w1 z1 c1 z21 z1 z2 zj 1j z1 @ j2
which shows that (A.2) and (A.3) are true for i 1. Assume that (A.2) and (A.3) are true for i 1, that is
A:8
212
2
c1
1 12
13
6 6 c2 1 23 3 6 1 12 6 z_1 6 6 7 6 13 1 23 6 7 6 6 76 4 5 6 6 6 zi 1 6 6 1;i 2 2;i 2 1 i 3;i 2 4 2
1;i 1
2;i 1 2 1;i1 1;i 2 3 6 6 2;i 2;i1 z1 6 6 7 6 6 7 6 6 7 6 4 5 6 6 6 zi 1 i 2;i1 4 i 2;i 1 i 1;i i 1;i1
1;i 1
1;i 2 2;i 2
1 i 2;i 2 ci 2
i 3;i 1
1n
7 7 7 7 7 7 7 7 7 i 2;i 1 7 7 7 1 i 2;i 1 7 5 ci 1 2;i 1
1 i 2;i 1
3
72 3 2n 7 7 zi 76 7 76 7 76 7 74 5 7 i 2;n 7 5 zn i 1;n
3
A:9
and V i 1
i 1 X k1
ck z2k zi 1 zi
n X
zj
ji
i 1 X
A:10
kj zk
k1
The zi -subsystem is i ' zi zi1 i
n X j 1 @ j1
i 1 i 1 X i 1 @ @ @t @e k k1
i zi1
@
'i zj
ek1
' k
n X @i j1
@
Gwj zj
n X j 1 @ j1
@
'k zj
i 1 n X i 1 X i 1 @ @ i T ek1 w ij zj @t @ek j1 k1
The derivative of Vi Vi 1 12 z2i is calculated as
n X @k j1
@
wj zj
!
A:11
V_ i
i 1 %
ck z2k zi zi1
k1
"
n %
zj
ji1
i 1 % k1
kj zk
n %
zj ij zi
ji1
i 1 %
i 1 i % i 1 % i 1 @ @ i ki zk ik zk zi zi 1 ek1 w i @t @ek k1 k1 k1
i 1 X
ck z2k zi zi1
k1
"
n X
ji1
zj
i X
213
#
kj zk
k1
i 1 i 1 X i 1 X i 1 @ @ i zi 1 ki ik zk ii zi ek1 w zi i @t @ek k1 k1
#
A12
From the de®nitions of ik , ik , ik and ik , it is easy to show that i as in (8.95) ki ik ki ik and that ii ii . Therefore the choice of results in (A.3) and zi
i 2 X k1
ki zk 1 i 1;i zi 1 ci zi 1 i;i1 zi1
n X
ki2
ik zk A:13
Combining (A.13) with (A.9), we get (A.2). We now rewrite the last equation of (8.94) as zn u ' n
n X j 1 @ j1
@
'n zj
n X @n j1
@
wj zj
n 1 n n X X n 1 X n 1 j 1 @ @ @ @k 'k zj wj zj ek1 ' k @t @e @ @ k j1 j1 k1
u
n 1 n X n 1 X n 1 @ @ nj zj ek1 w n @t @ek j1 k1
!
A:14
n follows the same de®nition as in (8.96). Noting that Va Vn 1 12 z2n where w and kn nk kn nk , and nn nn , (A.1) follows readily from (A.3) and (A.4).
References [1] Artstein, Z. (1983) `Stabilization with Relaxed Controls', Nonlinear Analysis, TMA±7, 1163±1173.
214
[2] Didinsky, G. and Bas° ar, T. (1994) `Minimax Adaptive Control of Uncertain Plants', Proceedings of the 33rd IEEE Conference on Decision and Control, 2839±2844, Lake Buena Vista, FL. [3] Freeman, R. A. and KokotovicÂ, P. V. (1996a) `Inverse Optimality in Robust Stabilization', SIAM Journal on Control and Optimization, 34, 1365±1391. [4] Freeman, R. A. and KokotovicÂ, P. V. (1996b) Robust Nonlinear Control Design. Birkhause, Boston. [5] Ioannou, P. A. and Sun, J. (1995) Robust Adaptive Control. Prentice-Hall, New Jersey. [6] KrsticÂ, M. (1996) `Invariant Manifolds and Asymptotic Properties of Adaptive Nonlinear Systems, IEEE Transactions on Automatic Control, 41, 817±829. [7] KrsticÂ, M. and KokotovicÂ, P. V. (1995) `Control Lyapunov Functions for Adaptive Nonlinear Stabilization', Systems & Control Letters, 26, 17±23. [8] Krstic M., Kanellakopoulos, I. and Kokotovic P. V (1992) `Adaptive Nonlinear Control Without Overparametrization', Systems & Control Letters, 19, 177±185. [9] Krstic M., Kanellakopoulos, I. and Kokotovic P. V. (1995) Nonlinear and Adaptive Control Design. Wiley, New York. [10] Li, Z. H. and Krstic M. (1996) `Geometric/Asymptotic Properties of Adaptive Nonlinear Systems with Partial Excitation, Proceedings of the 35th IEEE Conference on Decision and Control, 4683±4688, Kobe, Japan; also to appear in IEEE Transac. Autom. Contr. [11] Marino, R. and Tomei, P. (1995) Nonlinear Control Design: Geometric, Adaptive, and Robust. Prentice-Hall, London. [12] Pan, Z. and Bas° ar, T. (1996) `Adaptive Controller Design for Tracking and Disturbance Attenuation in Parametric-strict-feedback Nonlinear Systems', Proceedings of the 13th IFAC Congress of Automatic Control, vol. F, 323±328, San Francisco, CA. [13] Sepulchre, R., Jankovic, M. and Kokotovic P. V. (1997) Constructive Nonlinear Control. Springer-Verlag, New York. [14] Sontag, E. D. (1983) `A Lyapunov-like Characterization of Asymptotic Controllability', SIAM Journal of Control and Optimization, 21, 462±471. [15] Sontag, E. D. (1989) `A ``Universal'' Construction of Artstein's Theorem on Nonlinear Stabilization, Systems & Control Letters, 13, 117±123.
9
Stable adaptive systems in the presence of nonlinear parametrization A. M. Annaswamy and A.-P. Loh
Abstract This chapter addresses the problem of adaptive control when the unknown parameters occur nonlinearly in a dynamic system. The traditional approach used in linearly parametrized systems employs a gradient-search principle in estimating the unknown parameters. Such an approach is not sucient for nonlinearly parametrized (NLP) systems. Over the past two years, a new algorithm based on a min±max optimization scheme has been developed to address NLP adaptive systems. This algorithm is shown to result in globally stable closed loop systems when the states of the plant are accessible for measurement. We present the fundamental principles behind this approach and the stability properties of the resulting adaptive system in this chapter. Several examples from practical applications are presented which possess the NLP property. Simulation studies that illustrate the performance that results from the proposed algorithm and the improvement over the gradient scheme are also presented.
9.1
Introduction
The ®eld of adaptive control has, by and large, treated the control problem in the presence of parametric uncertainties with the assumption that the unknown parameters occur linearly [1]. This assumption has been fairly central in the development of adaptive estimation and control strategies, and commonly made in both discrete-time and continuous-time plants. Whether in an adaptive observer or an adaptive controller, the assumption of linear parametrization has dictated the choice of the structure of the dynamic system. For instance, in
216
adaptive observers, the focus has been on structures that allow measurable outputs to be expressed as linear, but unknown, combinations of accessible system variables. In direct adaptive control, a model-based controller is chosen so as to allow the desired closed loop output to be linear in the control parameters. In indirect adaptive control, estimators and controllers are often chosen so as to retain the linearity in the parameters being estimated. The design of stable adaptive systems using the classical augmented approach as in [1] or using adaptive nonlinear designs as in [2] relies heavily on linear parametrization. The problem is an important one both in linear and nonlinear dynamic systems, albeit for dierent reasons. In nonlinear dynamic systems, despite the fact that the majority of results have sought to extend the ideas of feedback linearization to systems with parametric uncertainties using the certainty equivalence principle, it is only within the context of linearly parametrizable nonlinear dynamics that global results have been available. Obviously, it is a nontrivial task to ®nd transform methods for general nonlinear systems so as to convert them into systems with linear parametrizations. In linear dynamic systems, it is quite possible to transform the problem into one where unknown parameters occur linearly. However, such a transformation also can result in a large dimension of the space of linear parameters. This has a variety of consequences. The ®rst is due to overparametrization which requires much larger amounts of persistent excitation or results in a lower degree of robustness. The second is that it can introduce undue restrictions in the allowable parametric uncertainty due to possible unstable pole-zero cancellations [3]. Nonlinearly parameterized systems are ubiquitous in nature, and as such unavoidable in many characterizations of observed complex behaviour. Friction dynamics [4], dynamics of magnetic bearing [5], and biochemical processes [6] are some examples where more than one physical parameter occurs nonlinearly in the underlying dynamic model. Several biological models of complex systems such as Hammerstein, Uryson, and Wiener models, consist of combinations of static nonlinearities and linear dynamic systems, which invariably result in NLP systems. The recent upsurge of interest in neural networks also stems from the fact that they can parsimoniously represent complex behaviour by including nonlinear parameterization. The fundamental property that linearly parametrized (LP) systems possess is in the cost function related to the parameter estimation. The latter can be posed as a hill-climbing problem, where the objective is to get to the bottom of the hill which is a measurable cost function of the parameter error. In LP systems, this hill is unimodal due to the linearity as a result of which, gradient rules suce to ensure convergence to the bottom of the hill which is guaranteed to be unimodal. In NLP systems, such a property is no longer preserved and hence, the gradient algorithm is not sucient to ensure either stability or
217
convergence. We show how algorithms that guarantee desired stability properties can be designed for general NLP systems in this chapter. Our proposed approach involves the introduction of a new adaptive law that utilizes a min±max optimization procedure, properties of concave functions and their covers, and a new error model. In the min±max optimization, the maximization is that of a tuning function over all possible values of the nonlinear parameter, and the minimization is over all possible sensitivity functions that can be used in the adaptive law. These two tuning functions lead to sign de®niteness of an underlying nonlinear function of the states and the parameter estimates. Such a property is then utilized to establish closed loop system stability. Concave/convex parametrization can be viewed as a special case of nonlinear parametrization. That is, if is an unknown parameter and is known to lie in a compact set , then the underlying nonlinearity is either concave with respect to all 2 or convex with respect to all 2 . We make use of properties of concave and convex functions in the development of the adaptive algorithms. To address adaptive control for general NP systems, a concave cover is utilized.The third feature of our approach is the development of an error model for NP systems. In any adaptive system, typically, there are two kinds of errors, which include a tuning error and parameter error. The goal of the adaptive system is to drive the former to zero while the latter is required to be at least bounded. Depending on the complexity of the dynamic system and the parametrization, the relationship between the two errors is correspondingly complex. A new error model is introduced in this chapter that is applicable to stable NLP adaptive system design. Very few results are available in the literature that address adaptive control in the presence of nonlinear parametrization [7, 8]. For these systems, the use of a gradient algorithm as in systems with linear parametrization will not only prove to be inadequate but can actually cause the system errors to increase. The approach used in [7], for example, was a gradient-like adaptive law which suces for stabilization and concave functions but would be inadequate either for tracking or for convex functions. In [9, 10, 11, 12], adaptive estimation and control of general NLP systems, both in continuous-time and discrete-time has been addressed, stability conditions for the closed loop system have been derived, and the resulting performance in various practical applications such as friction dynamics, magnetic bearing, and chemical reactors, has been addressed. An overview of the approach developed in these papers and the major highlights are presented in this chapter. In Section 9.2, we provide the statement of the problem, the applicability of the gradient rule, and the motivation behind the new algorithm. In Section 9.3, we present certain preliminaries that provide the framework for the proposed adaptive algorithms. These include de®nitions, properties of concave/convex functions, concave covers, solutions of min±max optimization problems, and sign de®niteness of related nonlinear functions. In Section 9.4, the new error
218
model is introduced, and its properties are derived. This is followed by stable adaptive system designs for both a controller and an observer. The ®rst addresses NLP systems when states are accessible, and the second addresses the problem of estimation in NLP systems when states are not accessible. Together, they set the stage for the general adaptive control problem when only the input and output are available for measurement. In Section 9.5, the performance of the proposed adaptive controller is explored in the context of physical examples, such as the position control of a single mass system subjected to nonlinear friction dynamics, positioning in magnetic bearings, and temperature regulation in chemical reactors. In Section 9.6, concluding remarks and extensions to discrete-time systems and systems where the matching conditions are not satis®ed are brie¯y discussed. Proofs of lemmas are presented in the appendix.
9.2
Statement of the problem
The class of plants that we shall consider in this chapter is of the form m X T _ fi
i
t; i
9:1 Xp Ap
pXp bp u '
t i1
n
where u is a scalar control input, Xp 2 R is the plant state assumed accessible for measurement, p; i 2 Rm , and 2 Rl are unknown parameters. i : R ! Rm , and ' : R ! Rl are known and bounded functions of the state variable. Ap is nonlinear in p, and fi is nonlinear in both i and i . Our goal is to ®nd an input u such that the closed loop system has globally bounded solutions and so that Xp tracks as closely as possible the state Xm of a reference model speci®ed as X_ m Am Xm bm r
9:2 where r is a bounded reference input. The plant in (9.1) has been studied in the literature extensively, whether or not the states of the system are accessible, but with the assumption that Ap is linear in p, and that fi is linear in i . Our focus here is when Ap and fi are nonlinear functions of the parameters p and i , respectively. This chapter is restricted to the case when all states of (9.1) are accessible. Further extensions to such systems with only input and output accessible are open problems and are currently under investigation. In order to evaluate the behaviour of the adaptive controller that is employed when the plant is linearly parametrized (which shall be referred to as an LP-
219
adaptive controller), we consider the simplest form of plants in (9.1), given by x_ p f
; u
9:3
where xp and are scalars, and is an unknown parameter and is a bounded function of xp . Choosing Am k < 0, and bm 1, the structure of the plant in equation (9.3) suggests that when is known, a control input of the form u
f
;
kxp r
9:4
leads to a closed loop system x_ p Am xp bm r and hence xp is bounded and tracks Xm asymptotically. The question is how u can be determined when is unknown and f is nonlinear in . The LP-adaptive control approach is to choose the input as u
De®ning ec xp xm ; ~ ^ equation is of the form of
^ f
;
kxp r
9:5
^ we obtain that the error ; and f^ f
; ,
e_c
kec f
f^
9:6
Suppose f is linear in with f
; g
then the gradient algorithm used in linear adaptive control approaches (see Chapter 3 in [1], for example) suggests that ~ must be adjusted as _ ~ ec rf^
9:7
The above adaptive law is derived from stability considerations by showing that V e2c ~2
9:8
is a Lyapunov function, which follows since the gradient of f with respect to ^ is g
and is independent of . When f is nonlinear in , the application of such a gradient approach is inadequate and in fact undesirable, since it can lead to an unstable behaviour. For instance, if V is chosen as in (9.8), the error equation (9.6) with the adaptive law as in (9.7) leads to a time derivative ~ ^ V_ ke2c ec f f^ rf
When f is concave, the property of a concave function enables us to conclude that V_ 0 when ec > 0; it also implies that ~ ^ V_ 0 if ec < 0 and kjec j < f f^ rf
220
This illustrates that, for a general f , it cannot be guaranteed that V_ 0 for all ec and ^ and thus can lead to unbounded solutions. To illustrate this, if we 2 choose f e , k 1 and 1, then for any which is such that j
tj 0 > 0;
8t
~ ^ > 0 for all we have rf^ < 0 and rf 1 if
t 0 ^ 2 1; p 20 1 ^ 2 p; 1 if
t 0 20
It follows therefore that when
t > 0, in the region
^ ^ ^ D
ec ; j 0 < kec < f f ; 2 1; we have
V_
since
ke2c ec f
kec < f
f^ < f
~ ^ >0 f^ rf
1 p 20
~ ^ f^ rf
Hence, D is an open invariant region where V_ > 0, since e_c > 0 at ec 0, f f^ 1 _ ^ ^ , and < 0 8 2 e_c < 0 for 0 < ec < 1; p .This implies that k 20 ^ the gradient algorithm will lead to
t ! 1 if
t > 0 and hence in ^ ! 1 if
t < 0. unbounded solutions. Similarly, we can show that
t This clearly illustrates the inadequacy of the gradient rule for nonlinear parametrizations. The discussion above indicates that the underlying problem is to ensure that the nonlinear function of the form ~ ec f f^ !
has to be made nonpositive by an appropriate choice of !. This has to be ensured independent of the sign of ec without introducing discontinuities in the controller. We illustrate how this can be accomplished in Section 9.4.
9.3
Preliminaries
The problem of NLP systems is addressed by making use of concavity/ convexity of the underlying nonlinearity wherever possible and by introducing concave covers in other cases. In this section, we introduce the basic de®nitions and properties of concave/convex functions as well as concave covers in
221
Sections 9.3.1 and 9.3.2, respectively.We show that these properties also ensure the desired sign-de®niteness property which will be used to design globally stable adaptive systems in Section 9.4.
9.3.1.1 De®nitions
De®nition
inequality
A function f
is said to be (i) convex on if it satis®es the
f
1
1
2 f
1
1
f
2 ;
81 ; 2 2
9:9
and (ii) concave if it satis®es the inequality f
1
1
2 f
1
1
f
2 ;
81 ; 2 2
9:10
where 0 1.
A simplex S in Rm is a convex polyhedron having exactly m 1 vertices. In order to obtain a continuous controller, we employ the following saturation function extensively in the control design:
The saturation function, sat
, is de®ned as y1
1 sat
y y jyj < 1
1 y 1
9.3.1.2 Properties of concave/convex functions
A useful property of these functions is their relation to the gradient. When f
is convex and dierentiable on , then it can be shown that f
f
0 rf0
0 ;
8; 0 2
9:11
and when f
is concave and dierentiable on , then f
where rf0
@f j . @ 0
f
0 rf0
0 ;
8; 0 2
9:12
222
Let J
!; f
;
^ !
^ f
;
a0 min max J
!; !2R
2
!0 arg min max J
!; !2R
fmax !0 max rf^
fmin min
9:14
9:15
2
where ^ 2 and is independent of . Then fmax fmin ^ f
min f^ min a0 max min 0
9:13
if f is convex on if f is concave on
if f is convex on
9:16
9:17
if f is concave on
where min ; max , fmax f
; max and fmin f
; min . Proof
See appendix.
For any x, and ; ^ 2 , ^ 0 a0 sat x 0; x f f^
! "
8 jxj > "
9:18
whether f is concave or convex, if a0 and !0 are chosen as in (9.16) and (9.17) respectively, with sign
x. Proof
See appendix.
When f is linear in , Lemma 3.2 is trivially satis®ed, which can be shown as follows. If f
; g
, then, a0 0 if !0 g
. With such an a0 , the lefthand side in (9.18) is identically zero. Lemma 3.2 implies that when f is nonlinear in , appropriate a0 and !0 can be found leading to an inequality as in (9.18) rather than an equality.
For a vector , let a0 minm max J
!;
9:19
!0 arg minm max J
!;
9:20
!2R
2S
!2R
2S
where J
!; is as in (9.13), ^ 2 S 2 and is a known nonzero constant. Then
A1 if f is convex on S
9:21 a0 0 if f is concave on S
!0
if f is convex on S if f is concave on S
A2 rf
where A A1 ; A2 T G 1 b,A1 is a scalar, A2 2 Rm 1
^ S1 T
f^ T ^ 1 ^
S2
f ; b G . .. . . .
f^ 1
^ Sm1 T
fS1 fS2 .. . fSm1
223
9:22
9:23
and fSi f
; Si . Proof
See appendix.
9.3.2.1 De®nitions
A point 0 2 c if 0 2 and rf0
0
f
f 0 ;
8 2
9:24
@f and f 0 f
; 0 . where rf0 X @ 0
c Xc \ , where c denotes the complement of c .
A concave cover of a function
f
f^ on is de®ned as:
F
f ij
f^;
ij
! c
8 2 c
8 2 ij 2 c
9:25
where !ij
fj j
fi ; i
cij f i
f^
!ij i ;
f i f
; i
9.3.2.2 Properties
F
as de®ned in (9.25) has the following properties:
(1) F
is concave on and F
f
f^; 8 2 .
9:26
224
(2) The solutions a0 and !0 for equations (9.14) and (9.15) are given by ^ a0 F
rf^ !0 !ij where !ij is de®ned as in (9.26). Proof
9:27 if ^ 2 c if ^ 2 ij 2 c
9:28
See appendix.
For any x, and ; ^ 2 ^ 0 x f f^
!
x a0 sat 0 "
9:29
by choosing a0 and !0 using (9.27) and (9.28) respectively. Proof 9.3.2.3
The proof follows along the same lines as that of Lemma 3.2. Examples of concave covers
To illustrate the nature and construction of a concave cover for a general function f , we present a number of examples in this section. We show that the nature of the cover not only depends on the function but also on the compact set that the parameter lies in.
Let
f
; g
cos
9:30
where g is a nonlinear function of the system variable . Obviously, if includes any interval whose length is larger than , f is neither concave nor convex on . We show how a concave cover F
can be constructed for dierent intervals of , assuming 1. (a) =2; 5=2: Since inequality (9.24) is satis®ed for all 0 in intervals 01 =2; 0 and 23 2; 5=2, we have that 01 ; 23 c . For any 0 2 12 0; 2, the inequality (9.24) is not satis®ed, and hence c 12 . At ®rst glance, one might be tempted to conclude that the interval =2; =2 2 c . This is not true since for any 0 2
0; =2, inequality (9.24) will not be satis®ed 8 2 since includes points beyond =2. Since 1 0 and 2 2, we have that !12 0
and
c12 1
f^
Hence, in this example, F
f f^ over 01 and 23 and is a straight line with zero gradient over 12 , as shown in Figure 9.1. Note that although F
depends on f^, the determination of the sets c and c in this example is independent of f^ since f^ only aects the vertical scale of F.
225
1 0.5 0 0.5 1 0
2
2
4
6
Figure 9.1
cos
. . . F
! Y =2; 5=2
(b) 1:4; 2: Unlike case (a), there are fewer subintervals in c in this case. We note that there does not exist any point 0 2 1:4; =2 which satis®es (9.24). Therefore, 1 1:4. Next, we need to ®nd a point 2 greater than =2 such that 2 satis®es (9.24). Since 2 lies on f f^ and 12 2 c , 2 is at the intersection of f f^ and the straight line joining f 1 and f 2 whose gradient is sin
2 at 2 . That is, 2 satis®es the relation cos
1:4
cos
2
2 1:4 sin
2
which yields 2 0:31365 rad. Proceeding as in the previous example, we note that 23 2 ; 0 2 c while 34 0; 2 2 c . The resulting F
is illustrated in Figure 9.2. It should be noted that 2 is close to but not coincident with the peak at zero. 1 0.5 0 0.5 1 5
0
5
Figure 9.2
cos
. . . F
! Y 1:4; 2
226
1 0.5 0 0.5 1 5
0
5
Figure 9.3
cos
. . . F
! Y 1:4; 1:4
(c) 1:4; 1:4: Case (b) implies that 12 1:4; 0:31365 2 c . Also, 0:313 65; 0 2 c. In this case, since includes only points up to 1:4, we also have that 0; 3 2 c for some 3 > 0. Proceeding as in case (b), we can show that 3 is the solution of the equation 1:4 sin
3 or 3 0:313 65. We also can conclude that 34 3 ; 1:4 2 c . The resulting F
as well as the convex cover which is computed in a similar fashion are shown in Figure 9.3. (d) ; ; n 0:5 ; n 2; n 0; 2; . . . ; even: We ®rst de®ne nx as the largest even integer that is smaller than x. Then c and c can be determined as c n 1 ; n ; n ; n 2 c ; n 1 ; n ; n ; n 2 ; cos
1:4
cos
3
3
where 1 and 2 are computed using the relations cos
cos 1 sin 1 1 sin 1
cos
cos 2 2 sin 2
sin 2
It is worthwhile noting that when ; n; n even, then c is an empty set and c n; n.
We note that f in (9.30) is independent of and as a result, c and c can be determined o-line which reduces the implementational burden of the proposed on-line algorithm. We next consider an example where c and c are
227
1 0.8 0.6 0.4 0.2 0 0.2
2
0
1
1
2
Figure 9.4 5
1 f
; 1e
. . . F
! Y 2; 2
functions of the system variable and thus cannot be determined easily online.
Consider f given by
1 f
; ; 2 2; 2 1 e 2; 1 where 1 is determined from In this case, c 1 ; 2 and c the following nonlinear equation :
2 1 e
1
1e
1
9:31
It is clear that the solution of equation (9.31) depends on . For a ®xed , a numerical routine such as the Newton±Raphson method could be used to solve for 1 . Figure 9.4 gives an illustration for the concave cover F
for 5. We note that F
always exist for any bounded f . Their complexities depend on the complexity of f . We also note that the number of subintervals in c (and c ) is proportional to the number of local maximas of f . Consequently, more computation is needed to determine all the i 's when f has several local maximas. In addition, if the i 's are independent of system variables, , then these i 's may be computed o-line. However, if they are dependent on , their on-line computation may prove to be too cumbersome. In the case of 2 Rm , the determination of the sets c and c generally involves a search over 2 such that
@f T j1
1
f f 1 ; 8 2 c 1 j @ Furthermore, given c and c , we need to de®ne hyperplanes, !T c such
228
that
F
!T c
f^ f ; 2 c We believe that the solutions to c , c ,! and c are nontrivial due to the dimensionality of the problem and is currently a subject of on-going research.
9.4
Stable adaptive NP-systems
We now address the design of stable adaptive systems in the presence of nonlinear parametrization. We address the problem of adaptive control of systems of the form of (9.1) in Section 9.4.2 with the assumption that states are accessible, propose a controller that includes tuning functions derived from the solution of a min±max optimization problem, and show that it guarantees global boundedness. In Section 9.4.3, we address the problem when the states cannot be measured, and propose a globally stable adaptive observer. We begin our discussion by proposing a new error model for NLP-systems in Section 9.4.1. " # $%&' All adaptive systems have, typically, two types of errors, a tracking error, ec , ~ that is not accessible for that can be measured, and a parameter error,, measurement. The nature of the relationship between these two errors is dependent on the complexity of the underlying plant dynamics as well as on the parametrization. This relationship in turn aects the choice of the adaptive algorithms needed for parameter adjustment. It therefore is important to understand the error model that describes this relationship and develop the requisite adaptive algorithm. In this section we discuss a speci®c error model that seems to arise in the context of NLP-adaptive systems. The error model for NLP-systems is of the form e c
9:32 e_c ec f f^ a0 sat " _
9:33 ^ e" !0 e c
9:34 e" ec " sat " ^ In (9.32), ec
t 2 R is the tracking error where f f
t; and f^ f
; . ~ while
t 2 R is the parameter error.
t 2 Rm is a given measurable function, while a0
t and !0
t are tuning functions to be determined.
^ 2 for all t t0 , and a0 and !0 are chosen as in (9.16) and
If
t (9.17) for concave±convex functions and (9.27) and (9.28) for nonconcave± convex functions, with sign
ec in all cases, then the solutions of the error ^ 2 8t t0 . model in (9.32)±(9.34) are globally bounded, provided
t
Proof
229
See Appendix.
^ 2 can be relaxed by making the following The requirement that
t changes in the adaptive laws in equation (9.33). _ e" !0
^ max
min
^ ;
>0
2 > max < min
9:33 0
This results in the following corollary.
^ 0 2 and ^ is adjusted as in (9.33 0 ), where a0 and !0 are If
t chosen as in (9.16) and (9.17) for concave±convex functions and (9.27) and (9.28) for nonconcave±convex functions, with sign
ec in all cases, then ec and ~ are always bounded. Proof
See appendix.
We note that unlike LP systems, the tracking error in (9.32) is a scalar. This raises the question as to how the error model can be applied for general adaptive system design where there is a vector of tracking errors. This is shown in the following section. " ( The following lemma gives conditions under which the stability properties of a vector of error equations can be reduced to those of a scalar equation.
Let
E_ Am E bv
9:35
e_c
9:36
kec v
If Am is asymptotically stable,
Am ; b controllable, and s eigenvalues of Am , then 9h such that for ec hT E:
k is one of the
(i) If ec is bounded, then E is bounded; and (ii) limt!1 ec
t 0 ) limt!1 E
t 0: Proof
See appendix.
" # Often, in NLP-systems, linear unknown parameters and multiplicative parameters are simultaneously present. This necessitates the need for the following
230
extension of equation (9.32): e_c
ec f
f^
~ 'T
~f^
ua
9:37
where is a constant whose sign and an upper bound jjmax are known. ~ 2 Rl are parameter errors, while '
t 2 Rl is a ~
t 2 R, ~ 2 R and
t measurable signal. In such a case, the following adaptive laws for adjusting ~ are needed, and the boundedness of the resulting errors is stated in ~, ~ and Lemma 8: _
9:38 ^ sign
e" !0 ^_ sign
e" f^
9:39
^_ sign
e" '
9:40
ec ua a0 jjmax sat "
9:41
^ 2 for all t t0 , and a0 and !0 are chosen as in (9.16) and
If
t (9.17) for concave±convex functions and (9.27) and (9.28) for nonconcave± convex functions, using sign
ec in all cases, then the solutions of the error model in (9.37)±(9.41) are bounded. Proof
See appendix.
^ 2 in Lemma 4.3 can be As in Corollary 4.1, the requirement that
t relaxed by making the following change in the adaptive law for ^ in equation (9.38). ^
>0 _ e" !0
;
2 ^
9:38 0 max max
min min
The corresponding result is stated in the following corollary:
^ 0 2 , and a0 and !0 are chosen as in (9.16) and (9.17) for If
t concave±convex functions and (9.27) and (9.28) for nonconcave±convex functions, using sign
ec in all cases, then the solutions of the error model in (9.37), (9.38 0 ), (9.39)±(9.41) are bounded. Proof 4.1.
Proof omitted since it follows closely that of Lemma 4.3 and Corollary
" The dynamic system under consideration here is of the form of equation (9.1), where u is a scalar control input, Xp 2 Rn is the plant state assumed accessible
231
for measurement, p; i 2 R, and 2 Rl are unknown parameters. i : R ! Rm , and ' : R ! Rl are known and bounded functions of the state variable. Ap is nonlinear in p, and fi is nonlinear in both i and i . Our goal is to ®nd an input u such that the closed loop system has globally bounded solutions and so that Xp tracks as closely as possible the state Xm of a reference model speci®ed in equation (9.2), where r is a bounded scalar input. We make the following assumptions regarding the plant and the model: (A1) Xp
t is accessible for measurement. (A2) i 2 i , where i X min;i ; max;i , and min;i and max;i are known; p is unknown and lies in a known interval P pmin ; pmax R. (A3)
t and '
t are known, bounded functions of the state variable Xp . (A4) f is a known bounded function of its arguments. (A5) All elements of A
p are known, continuous functions of the parameter p. (A6) bm bp where is an unknown scalar with a known sign and upper bound on its modulus, jjmax . (A7)
A
p; bp is controllable for all values of p 2 P, with A
p bm gT
p Am
where g is a known function of p. (A8) Am is an asymptotically stable matrix in Rn with det
sI
Am X Rm
s
s kR
s;
k>0
Except for assumption (A1), the others are satis®ed in most dynamic systems, and are made for the sake of analytical tractability. Assumption (A2) is needed due to the nonlinearity in the parametrization. Assumptions (A3)±(A5) are needed for analytical tractability. Assumptions (A6) and (A7) are matching conditions that need to be satis®ed in LP-adaptive control as well. (A8) can be satis®ed without loss of generality in the choice of the reference model, and is needed to obtain a scalar error model. Assumption (A1) is perhaps the most restrictive of all assumptions, and is made here to accomplish the ®rst step in the design of stable adaptive NLP-systems. Our objective is to construct the control input, u, so that the error, E Xp Xm , converges to zero asymptotically with the signals in the closed loop system remaining bounded. The structure of the dynamic system in equation (9.1) and assumptions (A6) and (A7) imply that when , p, , and i are known m fi
i ; i 'T
9:42 u
g
pT Xp r i1
meets our objective since it leads to a closed loop system X_ p Am Xp bm r
Our discussions in Section 9.2 indicate that an adaptive version of the
232
controller in (9.42), with the actual parameters replaced by their estimates together with a gradient-rule for the adaptive law, will not suce. We therefore propose the following adaptive controller: m ^ g
^ u pT Xp r fi
i ; ^i 'T ^ ua
t
9:43 i1
e" ec
" sat
_ ^ sign
^_
e c
9:44
"
e" ';
>0
9:45
> 0
9:46
^ sign
e" G;
_i sign
i e" !i 1
i i 2 i
i ^i max;i i > max;i
min;i i < min;i
^i ;
1 ; i > 0
9:47
p e" !m1 2
p p^ p p 2 P
p^ pmax p > pmax
pmin p < pmin
p_
2 ; p > 0
9:48
G
xp ; p g
pT Xp r; G^ g
^ pT Xp r ua
sign
sat
where ai min max sign
e" fi !i 2R i 2i
m1 e c ai " i1
f^i !i
^i
am1 jjmax min max sign
e" G^ !m1 2R
p2P
9:49
G
9:50
i ;
!m1
^ p
i 1; . . . ; m p
9:51
9:52
wi are the corresponding wi 's that realize the min-max solutions in (9.51) and (9.52), and jjmax denotes the maximum modulus of .The stability property of this adaptive system is given in Theorem 4.1 below. The closed loop adaptive system de®ned by the plant in (9.1), the reference model in (9.2) and (9.43)±(9.52) has globally bounded solutions if p^
t0 2 P and ^i
t0 2 i 8 i. In addition, limt!1 e"
t 0. Proof
For the plant model in (9.1), the reference model in (9.2) and the
233
control law in (9.43), we obtain the error dierential equation m b m ~G^ ua
t E_ Am E
G^ G fi
i ; i fi
i ; ^i 'T ~ i1
9:53
Assumptions (A6)±(A8) and Lemma 4.2 imply that there exists a h such that ec hT E and m 1 T~ ^ ^ ^ ~G ua
t fi
i ; i fi
i ; i ' e_c kec
G G i1
9:54
which is very similar to the error model in (9.37). De®ning the tuning error, e" , as in (9.44), we obtain that the control law in (9.43), together with the adaptive laws in (9.45)±(9.52) lead to the following Lyapunov function: m 1 2 1 1 1 ~2 p 1 p~2 e" ~T 1 ~ 1
1 ~2 V 2 jj jj jj i1 i i m 2 ^ ~ 2~ p
p p^
9:55 i
i i jj i1
This follows from Corollary 4.2 by showing that V_ 0. This leads to the ~ p~, ~ and ~i for i 1; . . . ; m. Hence ec is bounded global boundedness of e" , , and by Lemma 4.2, E is also bounded. As a result, i ; ' and the derivative e_c are bounded which, by Barbalat's lemma, implies that e" tends to zero. Theorem 4.1 assures stable adaptive control of NLP-systems of the form in (9.3) with convergence of the errors to within a desired precision ". The proof of boundedness follows using a key property of the proposed algorithm. This corresponds to that of the error model discussed in Section 9.4.1, which is given by Lemma 3.2. As mentioned earlier, Lemma 3.2 is trivially satis®ed in adaptive control of LP-systems, where the inequality reduces to an equality for !0 determined with a gradient-rule and a0 0. For NLP-systems, an inequality of the form of (9.18) needs to be satis®ed. This in turn necessitates the reduction of the vector error dierential equation in (9.53) to the scalar error dierential equation in (9.54). We note that the tuning functions ai and !i in the adaptive controller have to be chosen dierently depending on whether f is concave/convex or not, since they are dictated by the solutions to the min±max problems in (9.51) and (9.52). The concavity/convexity considerably simpli®es the structure of these tuning functions and is given by Lemma 3.1. For general nonlinear parametrizations,
234
the solutions depend on the concave cover, and can be determined using Lemma 4.3. Extensions of the result presented here are possible to the case when only a scalar output y is possible, and the transfer function from u to y has relative degree one [10].
" As mentioned earlier, the most restrictive assumption made to derive the stability result in Section 9.4.2 is (A1), where the states were assumed to be accessible. In order to relax assumption (A1), the structure of a suitable adaptive observer needs to be investigated. In this section, we provide an adaptive observer for estimating unknown parameters that occur nonlinearly when the states are not accessible. The dynamic system under consideration is of the form n b
psn 1 n i yp W
s; pu; W
s; p n i1
9:56 s i1 ai
psn i and the coecients ai
p and bi
p are general nonlinear functions of an unknown scalar p. We assume that
(A2-1) p lies in a known interval P pmin ; pmax . (A2-2) The plant in (9.56) is stable for all p 2 P. (A2-3) ai and bi are known continuous functions of p. It is well known [1] that the output of the plant, yp , in equation (9.56) satis®es a ®rst order equation given by y_ p
yp f
pT ;
>0
9:57
where !_ 1 !1 ku !_ 2 !2 kyp T
u; !T1 ; yp ; !T2 T f
p c0
p; c
pT ; d0 ; d
pT
9:58
9:59
9:60
9:61
for some functions c0
, c
, d0
,and d
, which are linearly related to bi
and ai
. in (9.58) and (9.59) is an
n 1
n 1 asymptotically stable matrix and
; k is controllable. Given the output description in equation (9.57), an obvious choice for an adaptive observer which will allow the on-line estimation of the nonlinear
235
parameter p, and hence ai 's and bi 's in (9.56), is given by y^_p
e 1 a0 sat "
^ yp f^T
where f^ f
^ p and e1 is the output error e1 y^p following error dierential equation can be derived: e_1
e1 f^
9:62 yp . It follows that the
e 1 a0 sat "
f T
9:63
Equation (9.63) is of the form of the error model in (9.32) with k1 , k2 1, '` ` 0, m 1, f1 f T , 1 , and 1 p. Therefore, the algorithm e" e1 p_
e 1 " sat "
p e" !0
p p^; p 2 P p p^ pmax p > pmax pmin p < pmin
p > 0
9:64
where a0 and !0 are the solutions of
a0 min max sign
e" !2R
p2P
f^
f
T
!0 arg min max sign
e" f^ !2R
p2P
f
T
!
^ p
p !
^ p
p
The stability properties are summarized in Theorem 4.2 below: For the linear system with nonlinear parametrization given in (9.56), under the assumptions (A2-1)±(A2-3), together with the identi®cation model in (9.62), the update law in (9.64) ensures that our parameter estimation problem has bounded errors in p~ if p^
t0 2 P. In addition, limt!1 e"
t 0. Proof The proof is omitted since it follows along the same lines as that of Theorem 4.1. We note that as in Section 9.4.2, the choices of a0 and !0 are dierent depending on the nature of f . When f is concave or convex, these functions are simpler and easier to compute, and are given by Lemma 3.1. For general nonlinear parametrizations, these solutions depend on the concave cover and as can be seen from Lemma 3.4, are more complex to determine.
236
9.5
Applications
) #' Friction models have been the focus of a number of studies from the time of Leonardo Da Vinci. Several parametric models have been suggested in the literature to quantify the nonlinear relationship between the dierent types of frictional force and velocities. One such model, proposed in [13] is of the form _
FS F FC sgn
x
_ FC sgn
xe
vx_s 2
Fv x_
9:65
where x is the angular position of the motor shaft, F is the frictional force, FC represents the Coulomb friction, FS stands for static friction, Fv is the viscous friction coecient, and vs is the Stribeck parameter. Another steady state friction model, proposed in [14] is of the form : _ F FC sgn
x
FC
_ S sgn
x
F
_ s 2 1
x=v
Fv x_
9:66
Equations (9.65) and (9.66) show that while the parameters FC ; FS and Fv appear linearly, vs appears nonlinearly. As pointed out in [14], these parameters, including vs , depend on a number of operating conditions such as lubricant viscosity, contact geometry, surface ®nish and material properties. Frictional loading, usage, and environmental conditions introduce uncertainties in these parameters, and as a result these parameters have to be estimated. This naturally motivates adaptive control in the presence of linear and nonlinear parametrization. The algorithm suggested in Section 9.4.2 in this chapter is therefore apt for the adaptive control of machines with such nonlinear friction dynamics. In this section, we consider position control of a single mass system in the presence of frictional force F modelled as in equation (9.65). The underlying equations of motion can be written as x F u
9:67
where u is the control torque to be determined. A similar procedure to what follows can be adopted for the model in equation (9.66) as well. Denoting _ x _ T; ' sgn
x;
1=v2s ;
_ sgn
xe _ f
x; FS
x_ 2
;
FC
FC ; Fv T
9:68
it follows that the plant model is of the form _ 'T u x f
x;
9:69
237
_ is convex for all x_ > 0 and concave for all x_ < 0. We choose a where f
x; reference model as 2 s 2!n s !2n xm !2n r
9:70
where and !n are positive values suitably chosen for the application problem. It therefore follows that a control input given by ^ a0 sat ec u kec D1
sx !2n r 'T ^ ^f
;
9:71 " where D1
s 2!n s ! 2n , together with the adaptive laws _ ~
>0
9:72
_ ~ e" !0 ;
> 0
9:73
~_ e" f^;
> 0
9:74
e" ';
with a0 and !0 corresponding to the min±max solutions when sign
e" f is concave/convex, suce to establish asymptotic tracking. We now illustrate through numerical simulations the performance that can be obtained using such an adaptive controller. We also compare its performance with other linear adaptive and nonlinear ®xed controllers. In all the simulations, the actual values of the parameters were chosen to be FC 1 N;
FS 1:5 N;
Fv 0:4 Ns/m;
vs 0:018 m/s
9:75
and the adaptive gains were set to
diag
1; 2;
2;
108
9:76
The reference model was chosen as in equation (9.70) with = 0.707, !n = 5 rad/s, r sin
0:2t. Simulation 1 We ®rst simulated the closed loop system with our proposed controller. That is, the control input was chosen as in equation (9.71) with ^ k 1, and adaptive laws as in equations (9.72)±(9.74) with " = 0.0001.
0 was set to 1370 corresponding to an initial estimate of ^vs = 0.027 m/s, which is 50% larger than the actual value. Figure 9.5 illustrates the tracking error, e x xm ,the control input, u, and the error in the frictional force, ^ where F is given by (9.65) and F^ is computed from (9.65) by eF F F, replacing the true parameters with the estimated values. e, u and eF are displayed both over [0, 6 min] and [214 min, 220 min] to illustrate the nature of the convergence. We note that the position error converges to about 5 10 5 rad, which is of the order of ", and eF to about 5 10 3 N. The discontinuity in u is due to the signum function in f in (9.68).
(rad)
238
(N)
(a)
* (N)
(b)
(c) Figure 9.5 $ + ,- e . ,- u . ,- eF
Simulation 2 To better evaluate our controller's performance, we simulated another adaptive controller where the Stribeck eect is entirely neglected in the friction compensation. That is
so that the control input u
_ F^v x_ F^ F^C sgn
x kec
D1
sx !2n r
9:77 F^
9:78
with estimates F^C and F^v obtained using the linear adaptive laws as in [1]. As before, the variables e, u and eF are shown in Figure 9.6 for the ®rst 6 minutes as well as for T [214 min, 220 min]. As can be observed, the maximum position error does not decrease beyond 0.01 rad. It is worth noting that the control input in Figure 9.6 is similar to that in Figure 9.5 and of comparable
239
(rad)
(N)
(a)
* (N)
(b)
(c) Figure 9.6 % # / 0 + ,- e . ,- u . ,- eF
magnitude showing that our min±max algorithm does not have any discontinuities nor is it of a `high gain' nature. Note also that the error, eF does not decrease beyond 0.5 N. Simulation 3 In an attempt to avoid estimating the nonlinear parameters vs , in [15], a friction model which is linear-in-the-parameters was proposed. The frictional force is estimated in [15] as _ F^S jxj _ 1=2 sgn
x _ F^v x_ F^ F^C sgn
x
9:79
with the argument that the square-root-velocity term can be used to closely match the friction-velocity curves and linear adaptive estimation methods similar to Simulation 2 were used to derive closed loop control. The resulting
(rad)
240
(N)
(a)
* (N)
(b)
(c) Figure 9.7 % # ,2- ,- e . ,- u . ,- eF
performance using such a friction estimate and the control input in equation (9.78) is shown in Figure 9.7 which illustrates the system variables e, u and eF for T [0, 6 min] and for T [360 min, 366 min]. Though the tracking error remains bounded, its magnitude is much larger than those in Figure 9.5 obtained using our controller. ) / ,1Stirred tank reactors (STRs) are liquid medium chemical reactors of constant volume which are continuously stirred. Stirring drives the reactor medium to a uniform concentration of reactants, products and temperature. The stabilization of STRs to a ®xed operating temperature proves to be dicult because a
241
few physical parameters of the chemical reaction can dramatically alter the reaction dynamics. De®ning X1 and X2 as the concentration of reactant and product in the in¯ow, respectively, r as the reaction rate, T as the temperature, h as the reaction heat released during an exothermic reaction, d as the volumetric ¯ow into the tank, T T Tamb , it can be shown that three dierent energy exchanges aect the dynamics of X1 [16]: (i) conductive heat loss with the environment at ambient temperature, Tamb , with a thermal heat transfer coecient, e,(ii) temperature dierences between the in¯ow and out¯ow which are Tamb and T respectively, (iii) a heat input, u, which acts as a control input and allows the addition of more heat into the system. This leads to a dynamic model 1 X_1 0 exp X1 d
X1in X1 T Tamb 1
9:80 X1 dX2 X_2 0 exp T Tamb q 1 1 _ T h0 exp T X1 u T Tamb The thermal input, u, is the only control input. (Volumetric feed rate, d, and in¯ow reactant concentration, X1in are held constant.) To drive T from zero (Tamb ) to the operating temperature, Toper , we can state the problem as the tracking of the output Tm of a ®rst order model, speci®ed as T_m
kTm kToper
9:81
where k > 0. Driving and regulating an STR to an operating temperature is confounded by uncertainties in the reaction kinetics. Speci®cally, a poor knowledge of the constants, 0 and 1 , in Arrhenius' law, the reaction heat, h, and thermal heat transfer coecient, e, makes accurately predicting reaction rates nearly impossible. To overcome this problem, an adaptive controller where 0 , 1 , h and e are unknown may be necessary. ) The applicability of the adaptive controller discussed in Section 9.4.2 becomes apparent with the following de®nitions 1 1 0 ; 2 1 ; ; Ap q=; bp 1=; 0 2 2 f 1 h exp X1 ; Te T Tm ; T T Tamb which indicates that equation (9.80) is of the form (9.1). Note that f is a convex
242
function of . (It is linear in 1 0 and exponential in 2 1 . If reaction heat, h, is unknown, it may be incorporated in 1 , i.e. 1 h0 .) Since 0 , 1 and h are unknown constants within known bounds, assumption (A2) is satis®ed. Assumption (A1) is satis®ed since the temperatures are measurable. The system state, T , complies with (A3). Furthermore f is smooth and dierentiable with known bounds and hence (A4) is satis®ed. Finally, (A6)± q . Since e is (A8) are met due to the choice of the model as in (9.81) for k unknown, is unknown and is therefore estimated. Since the plant is ®rst order, the composite error ec is given by ec Te . Referring to the adaptive controller outlined in Section 9.4.2, the control input and adaptation laws are as follows: e c ^ T ^ kToper u f
; a sat " e c e" ec "sat ; ">0 " ^_ e" T ; >0 _ ^ ; ^ e" !
> 0
if i 2 i;min ; i;max ^i
^i i;min if i i;min
i;max if i i;max Simulation results have shown that our proposed controller performs better than a linear adaptive system based on linear approximations of the nonlinear plant dynamics. Due to space constraints, however, the results are not shown in this chapter but the reader may refer to [11] for more details.
) 3+ + Magnetic bearings are currently used in various applications such as machine tool spindle, turbo machinery, robotic devices, and many other contact-free actuators. Such bearings have been observed to be considerably superior to mechanical bearings in many of these applications. The fact that the underlying electromagnetic ®elds are highly nonlinear with open loop unstable poses a challenging problem in dynamic modelling, analysis and control. As a result, controllers based on linearized dynamic models may not be suitable for applications where high rotational speed during the operation is desired. Yet another feature in magnetic bearings is the fact that the air gap, which is an underlying physical parameter, appears nonlinearly in the dynamic model. Due
243
to thermal expansion eects, there are uncertainties associated with this parameter. The fact that dynamic models of magnetic bearings include nonlinear dynamics as well as nonlinear parametrizations suggests that an adaptive controller is needed which employs prior knowledge about these nonlinearities and uses an appropriate estimation scheme for the unknown nonlinear parameters. To illustrate the presence of nonlinear parametrization, we focus on a speci®c system which employs magnetic bearings which is a magnetically levitated turbo pump [5]. The rotor is spun through an electric motor, and to actively position the rotor, a bias current i0 is applied to both upper and lower magnets and an input u is to be determined by the control strategy. For a magnetic bearing system where rotor mass is M operating in a gravity ®eld g, the rotor dynamics is represented by a second order dierential equation of the form n2 0 A
i0 0:5u2 n2 0 A
i0 0:5u2 M z Mg
9:82 4
h0 z2 4
h0 z2 where n denotes the number of coils, 0 the air permeability, A the pole face area, i0 the bias current (with juj < 2i0 ), h0 the nominal air gap, and z the rotor position. One can rewrite equation (9.82) as z
g f1
h0 ; ; z f2
h0 ; ; zu f3
h0 ; ; zu2 ;
where f1
h0 ; ; z f2
h0 ; ; z f3
h0 ; ; z
juj < 2i0
9:83
n2 0 A 4M 4h0 zi02
h0
z2
h0 z2
h0
z2
h0 z2
h0
z2
h0 z2
2
h20 z2 i0 h0 z
zi02 1
h0 ; z i0 2
h0 ; z z 3
h0 ; z
The control objective is to track the rotor position with a stable second order model as represented by the following dierential equation: zm c1 z_m c2 zm r
9:84
) By examining equation (9.83), it is apparent that the parameter h0 occurs nonlinearly while occurs linearly. An examination of the functions f1 , f2 u, and f3 u2 further reveals their concavity/convexity property and are
244
Table 9.1
& f1 ; f2 u f3 u 2 hi
Function
Concavity/convexity
Monotonic property
Prerequisite
F 1 f1
convex concave convex concave convex concave
decreasing increasing decreasing increasing decreasing increasing
0 < z < hmin hmin < z < 0 u>0 u0
9:89
i > 0
9:90
i
and the linear parameters as ^_1 e" 1 ^1 zi02 ;
^_2 e" 2 ^2 i0 u;
^_3 e" 3 ^3 zu2 ;
where i are positive. Since the functions Fi are either convex or concave, ai and wi are chosen as follows: (a) Fi is convex sat
ec ="max Fimax ai 0
F^i
Fimax hmax
Fimin ^
h hmin
hmin
if e" 0 otherwise
Fimax Fimin
sat
ec =" h hmin max !i
t @F
sat
ec =" i @h h^
(b) Fi is concave 0 ai sat
ec ="max F^i @Fi
sat
ec =" @h h^ !i
t
sat
ec =" Fimax hmax
Fimin
Fimin hmin
245
if e" 0 otherwise
Fimax hmax
Fimin ^
h hmin
hmin
if e" 0 otherwise
if e" 0 otherwise
By examining equation (9.85), it is apparent that f^2 cannot be arbitrarily small. This requirement is satis®ed by disabling the adaptation when the magnitude of f^2 reaches a certain threshold. The adaptive controller de®ned by equations (9.85)±(9.90) guarantees the stability of the magnetic bearing system as shown by the simulations results in [11].
9.6
Conclusions
In this chapter we have addressed the adaptive control problem when unknown parameters occur nonlinearly. We have shown that the traditional gradient algorithm fails to stabilize the system in such a case and that new solutions are needed. We present an adaptive controller that achieves global stabilization by incorporating two tuning functions which are selected by solving a min±max optimization scheme. It is shown that this can be accomplished on-line by providing closed-form solutions for the optimization problem. The forms of these tuning functions are simple when the underlying parametrization is concave or convex for all values of the unknown parameter and more complex when the parametrization is a general one. The proposed approach is applicable to discrete-time systems as well. How stable adaptive estimation can be carried out for NLP-systems has been addressed in [12]. Unlike the manner in which the tuning functions are introduced in continuous-time systems considered in this chapter, in discretetime systems, the tuning function a0 is not included in the control input, but takes the form of a variable step-size t in the adaptive-law itself. It is once again shown that a min±max procedure can be used to determine t as well as the sensitivity function !t at each time instant. As in the solutions presented here, !t coincides with the gradient algorithm for half of the error space.
246
The class of NLP adaptive systems that was addressed in this chapter is of the form of (9.1). It can be seen that one of the striking features of this class is that it satis®es the matching conditions [17]. Our preliminary investigations [18] show that this can be relaxed as well by judicious choice of the composite scalar error in the system, thereby expanding the class of NLP adaptive systems that are globally stabilizable to include all systems that have a triangular structure [19].
Appendix
Proof of lemmas
& %
We establish (9.14) and (9.15) by considering two cases: (a) f is convex, and (b) f is concave, separately. (a) f is convex. Since ! is linear in , in this case, the function J
!; given by J
!; f f^ !
^
A:1 is convex in . Therefore, J
!; attains its maximum at either min or max or both. The above optimization problem then becomes ! min max fmin f^ !
^ min ; fmax f^ !
^ max
A:2 !2R
or equivalently it can be converted to a constrained linear programming problem as follows: min z
!;z2R2
subject to
fmin fmax
f^ !
^ f^ !
^
min z max z
A:3
By adding slack variables "1 0 and "2 0, the inequality constraints in (A.3) can be further converted into equality constraints fmin f^ !
^ min "1 z
A:4 fmax f^ !
^ max "2 z
A:5 Solving for ! in equation (A.4) and substituting into equation (A.5), we
247
have z
fmin
max ^ max min
fmax
min ^ max min
f^
"1
max ^ "2
^ max min max
min min
A:6
The optimal solution can now be derived by considering three distinct cases: (i) min < ^ < max , (ii) ^ max and (iii) ^ min .
(i) min < ^ < max : Since the last two terms in equation (A.6) are positive for "1 0 and "2 0, minimum z is attained when "1 "2 0 and is given by fmax fmin ^
min
A:7 zopt fmin f^ max min The corresponding optimal ! can be determined by substituting equation (A.7) in equation (A.4) and is given by !opt
fmax max
fmin min
A:8
(ii) ^ max : equation (A.6) can be simpli®ed as z "2
and thus minimum z is obtained when "2 0 or zopt 0. The corresponding optimal ! using equation (A.4) is given by !opt
fmax fmin "1
max min
A:9
Thus, !opt is nonunique. However, for simplicity and continuity, if we choose "1 = 0, !opt is once again given by equation (A.8). (iii) ^ min : Here equation (A.6) reduces to z "1
Hence minimum z corresponds to "1 0 or zopt 0. Using equation (A.5), it follows that !opt is given by !opt
fmax fmin "2
max min
A:10
Again, !opt is nonunique but can be made equal to equation (A.8) by choosing "2 = 0. (b) f is concave. Here, we prove Lemma 3.1 by showing that (i) the absolute minimum value for a0 0, and (ii) this value can be realized when ! !0 as in (9.17). (i) For any !, we have that
248
f
f^ !
^
since ^ 2 . Hence, for any !
0;
for some 2
max J
!; 0 2
and as a result
min max J
!; 0 !2R
2
Hence, a0 attains the absolute minimum of zero for some and some !. (ii) We now show that when ! !0 , a0 0, where !0 is given by (9.17). Since f is concave, J is concave as well. Using the concavity property in (9.12), we have f f^ rf^
^ 0 That is,
max f
f^ rf^
^
2
0
A:11
Equation (A.11) implies that if we choose ! rf^, then min max J
!; 0 !2R
2
which proves Lemma 3.1. & %
We note that a0 sign
x sat
x a0 ; "
8jxj > "
by de®nition of the sat
function. If a0 and !0 are chosen as in (9.16) and (9.17) respectively, then a0 max f f^ !0
^ 2
which implies that for any 2 , ^ 0 sign
x f f^
!
a0 sign
x sat
since sign
x. Inequality (9.18) therefore follows.
x 0 "
& %
When f is convex, J
!; is also convex and the min±max problem minm max J
!; minm max f f^ !T
^ !2R
2S
!2R
2S
A:12
can be converted into a constrained LP problem involving the
m 1 vertices
249
of the simplex S . De®ning max2S J
!; z, (A.12) can now be expressed as: min z
z;!2Rm1
A:13 subject to: g
!; Si J
!; Si z 0; i 1; . . . ;
m 1 We solve (A.13) by converting into an unconstrained problem as follows. Rewriting the constraints in matrix form as H
x Gx
where x z
!T and 1
^ 1
^ G . . . 1
^
S1 T
b0
f^
f^ b
f^
S2 T ; .. . Sm1 T
fS1
fS2 .. . fSm1
A:14
we have that rx H
x G and G is full rank since Si are distinct vertices of S and is nonzero. De®ning the Lagrangian function by
x; z T H
x
A:15
where 1 ; 2 ; . . . ; m1 T , and i ; i 1; . . . ; m 1 are the Lagrange multipliers, the Kuhn Tucker theorem states that m1 1 i1 i rx
x ; m1 0
A:16 i
^ Si i1
r
x ;
Gx T
T
b 0
A:17
where x is the optimal solution. From (A.16), we have m i1
i
Sm1
Si Sm1
^
A:18
Three cases of ^ will now be considered: (a) ^ is in the interior of S , (b) ^ is on the boundary of S , (c) ^ coincides with one of the vertices, Si ; i 1; . . . ; m 1. Case (a)
This implies that ^
m 1 i1
i Si ; with
m1 i1
i 1; 0 < i < 1
250
Substituting into equation (A.18), we have m i1
i
i
Sm1
Si 0
m independent vectors, it follows that Since
Sm1 Si ; i 1; ; m are m i i 0 or i i ; m1 1 i1 i . Thus i > 0 for all i. Therefore, from equation (A.17), we require Gx b 0 and thus the optimal solution is given by x G 1 b or a0
G 1 b11 where
Aij refers to the ijth element of a matrix A. Case (b) ^ is on the boundary of S . In this case, ^ is a linear combination of at most m vertices. Suppose ^
r
i Si ;
r i1
i1
i 1; 0 < i < 1; 1 r m
where, for convenience, we have reordered the vertices such that ^ is a linear combination of the ®rst r vertices. From (A.18), we have that r i1
i
i
Sm1
Si 0
Thus, i i ; i 1; . . . ; r and r1 r2 . . . m 0. Equation (A.17) requires that r i
fSi f^
^ Si T ! z 0 i1
) since
r
i1
r i1
i fSi
f^
^
r i1
i Si T !
z0
i 1. Since the term within the parentheses is zero, we have that z a0
r i1
i fSi
f^
A:19
While equation (A.19) gives a nice closed form solution, this optimal solution is not so readily computable because we require the values of the Lagrange multipliers which in turn can only be computed by decomposing the estimate ^ in terms of the m 1 vertices. We avoid this that the optimal # by showing $ solution in (A.19) coincides onces again with G 1 b 11 . Let G be partitioned as 1
^ S1 T
f^ fS1 G ; b Em1 Amm Bm1
Then G
where M E
^ f^ a0 .
1
2
1 ME
3
r 0 0
M
1
S1 A T
and
G 1 b11
hence
251
r
i1
i fSi
Case (c) ^ coincides with one of the vertices, Sj , for some 1 j m 1: From (A.18), j j 1; i i 0; 8i 6 j; 1 i m 1. In order to satisfy equation (A.17), we require that fSj
f^
^
Sj T !
z0
A:20
Since ^ Sj , (A.20) implies that the optimal solution is z a0 0. We can # $ show that a0 G 1 b 11 in this case as well as in what follows. Rewriting G such that the jth constraint corresponding to ^ Sj is the ®rst constraint, we have that 1 01m 0 G ; b Em1 Amm Bm1 and hence its inverse is given by G
1
1
01m
A 1E
A
1
Hence,
G 1 b11 0 a0 . The proof for the case when f is concave is the same as part (b) in the proof of Lemma 3.1. & %
" (1) For 2 c , F
f f^ and is concave since f is concave on c . For 2 c , F
is linear in and thus is also concave. At i ; i 0; 1; . . . ; n 1, F
i
f i f^. Thus F
is a continuous concave function on . In addition, for 2 ij , F
f j
1
f i
f^;
j
i i
Since 0 1 and f is not concave on each ij , it follows that f
f j
1 f i for all in each ij and hence F
f (2) We ®rst consider the expression J
!; f
f^; 8 2 c f^ !
^
252
For any , ! 2 R and ^ 2 , there exists some 2 such that J 0, implying that min max J
!; 0:
A:21 !2R 2
(A.21) implies that the absolute minimum for the min±max problem is zero. Now, if
f
f^
! c;
then we have that
for some ! and c
J
!;
!^ c
Hence, for some ! and c, max J
!;
!^ c 2
Therefore, min max J
!; !2R 2
subject to
min!;c2R
!^ c
f f^
! c
A:22
We have thus converted the min±max problem into a constrained linearproblem in (A.22). We can now establish (9.27) by considering the equivalent problem in (A.22) for two cases of : (a) ^ 2 c , and (b) ^ 2 c . (a) ^ 2 c : By the de®nition of c , we have that rf^
^
f
f^;
8 2
A:23
which satis®es the constraints in (A.22) if we choose ! rf^ and ^ For these choices of ! and c, it follows that c rf^. min
!^ c 0
!;c2R
^ 0 and hence the min±max problems in (9.14) and (9.15) Since ^ 2 c , F
have solutions @f ^ a0 F
0; !0 @ ^
(b) ^ 2 c : Suppose
^ 2 ij for some i; j
From (A.22), we have min max f f^ !
^ !2R
2
s.t. s.t.
f
min!;c
!^ c f^
! c; 8 2
min!;c
!^ c
F
! c;
8 2
A:24
A:25
A:26
since F
f
253
f^. However, since F
is concave on , we have
F
rF1
1 F
1 ;
8; 1 2
A:27
Since from (9.25), F
!kl ckl ;
8 2 kl
A:28
for any 1 2 kl , it follows that rF1 !kl . Therefore, (A.27) can be rewritten as F
!kl ckl ;
8 2
A:29
8 2 = kl
A:30
(A.28) and (A.29) imply that F
<
!kl ckl ;
since the intervals kl 8k; l are unique and equality only holds for the interval kl . Therefore, the optimization problem in (A.26) can be further transformed as follows: min!;c
!^ c min!;c
!^ c s.t. F
!kl ckl ; 2 kl s.t. F
! c; 8 2 F
<
!kl ckl ; = 2kl
A:31 Since the active constraints on the right-hand side of (A.31) occur only in the set kl , the optimal solution of (A.31) is simply given by
!kl ^ ckl . We note that the expansion of the constraint in (A.26) into constraints in (A.31) is not unique since the choice of k and l are arbitrary. Hence, the optimal solution of (A.26) has to be the minimum of all possible solutions of (A.26) derived from considering all the possible sets of constraints that can be derived from the single constraint in (A.26). In other words
s.t.
min!;c
!^ c min
!kl ^ ckl ; F
! c; 8 2
k; l 1; . . . ; n
!
A:32
With i; j chosen as in (A.24), using equation (A.28), we obtain that ^
!ij ^ cij F
^2kl , inequality (A.30) implies that since ^ 2 ij .When i 6 k; j 6 l, since = ^ <
!kl ^ ckl ; F
8k 6 i; l 6 j:
That is, the minimum solution in (A.32) occurs when k i; l j and
254
therefore, min!;c
!^ c
F
! c
s.t.
^
!ij ^ cij F
A:33
The corresponding optimal ! is given by !0 !ij . We next show how the inequality in (A.26) is actually attained with equality by establishing that if the concave cover F
was constructed such that it is strictly greater than
f f^, then the corresponding optimal solution will be larger. On the other hand, if F
was constructed such that it is less than
f f^ for some , the corresponding optimal solution will be smaller. Hence if F
f f^, the two solutions will be equal. This is established formally below. From (A.26) and (A.33), we have that min!;c
!^ c s.t.
f
f^
! c
min!;c
!^ c s.t. F
! c
!ij ^ cij :
Suppose we perturb F
as F 0
F
"
f min!;c
!^ c
s.t.
f
f^
! c
Let c0 c
f^ "; " > 0, then
min!;c
!^ c
s.t.
A:34
F 0
! c min!;c
!^ c
s.t.
F
! c
"
A:35
"= in (A.35). Then the optimal solution of (A.35) is given by s.t.
min!;c0
!^ c0 " F
! c0
!ij ^ cij "
A:36
following the result in (A.33). On the other hand, suppose that F 0
F
". This implies that 0 F
f f^ for some . Using the same arguments as in (A.34)± (A.36), we obtain that min!;c
!^ c min!;c
!^ c
!ij ^ cij ^ s.t. F 0
! c s.t.
f f
! c
":
Hence, we have
!ij ^ cij
"
min!;c
!^ c
!ij ^ cij " ^ s.t.
f f
! c
A:37
255
Since the concave cover, F
, was constructed as a tight bound over
f f^ i.e. F
f f^, we have that " 0 in (A.37), and hence min!;c
!^ c
s.t.
f^
! c
f
!ij ^ cij ^ F
A:38
for a corresponding optimal ! given by !0 !ij . Therefore, statement (2) in Lemma 3.4 holds. & %
" Suppose we choose a Lyapunov candidate given by
where ~ ^
V 12
e2" ~2
A:39
. Then taking the derivative of V with respect to time yields _ V_ e" e_" ~~
e2" .
A:40
Let y Since the discontinuity at jec j " is of the ®rst kind and since e" 0 when jec j ", it follows that the derivative V_ exists for all ec , and is given by V_ 0
when jec j "
When jec j > ", substituting (9.32) and (9.33) into (A.40), we have ~ 0 a0 sat ec V_ e" ec e" f f^ ! "
Equation (A.42) can be simpli®ed, by the choice of e" as ~ 0 a0 sat ec V_ e2" e" f f^ ! "
A:41
A:42
A:43
If a0 and !0 are chosen as in (9.16) and (9.17) for concave-convex functions and (9.27) and (9.28) for nonconcave±convex functions, from Lemmas 3.2 and 3.5, it follows that ~ 0 a0 sat ec 0 ec f f^ ! " Since sign
ec sign
e" ; 8 jec j > " it follows that
e" f
~ 0 f^ !
As a result, equation (A.43) reduces to V_
e2" 0;
a0 sat
e c 0 "
for jec j "
A:44
(A.41) and (A.44) imply that V is indeed a Lyapunov function which leads to
256
From the de®nition of e" , it follows that ec is global boundedness of e" and . also bounded. & " To account for the projection terms in equation (9.3 ), we modify our Lyapunov function candidate as ^
A:45 V 12 e2" 2 2
it follows that V is positive de®nite and radially From the de®nition of ^ and , unbounded with respect to e" , ~ and [20]. Equations (A.45) and (9.33 0 ) yield a time derivative ~ a0 sat ec m
; ^ V_ e2" e" f f^ ! " where ^
^
^ 1
^_ ^ m
;
then we can obtain an expression for V_ similar to If m 0 for all ^ and , (A.43). We show this by considering the following three cases: (a) 2 , (b) > max , (c) < min . In case (a), ^ and hence m 0. In both cases _ (b) and (c), ^ 0, and by the choice of ^ and and since 2 , it follows that m 0. As a result, ~ 0 a0 sat ec V_ e2" e" f f^ ! " 0 ^ Also, the algorithm in equation (9.33 ) ensures that
t 2 for all t t0 . Therefore, the same arguments following equation (A.43) in the proof of Lemma 4.1 can be used to conclude that V_ 0;
8ec
Therefore, V is a Lyapunov function, which leads to the boundedness of e" ; ~ [20]. By the de®nition of e" , ec is also bounded, thus proving Corollary and , 4.1. & %
" Let the characteristic polynomial of Am be a
s
s kR
s where R
s is a Hurwitz polynomial. By the choice of Am , we have that a
s
sI
Am 1 b P
s
A:46
where P
s is an n 1 matrix whose elements are polynomials with
257
degrees n 1. Denoting S 1; s; ; sn 1 T , we have, from the controllability of
Am ; b that P
s MS where M is nonsingular We choose h as h MT
1
A:47
p
where p is an n-dimensional vector such that pT S R
s. Therefore, a
shT
sI
Am 1 b hT MS R
s
using equation (A.47)
As a result, we have hT
sI
Am 1 b
1 sk
A:48
and it follows by the choice of ec that e_c
kec v
(i) Since R
s is a Hurwitz polynomial, we have that si ec 2 L 1 R
s Since from (A.48)
we have that
if ec 2 L 1
1 1 v ec a
s R
s si v 2 L1 a
s
A:49
if ec 2 L 1
and therefore, since
Am ; b is controllable, it follows that E 2 L 1 . (ii) follows using the same arguments. & %
" Let the Lyapunov candidate be chosen as ~2 V 12
e2" jj~2 jj~ 2 jj For jec j ", V_ 0, and for jec j > ", ~ 0 _ V e" jj sign
f f^ !
jjmax ec a0 sat " jj
With a0 , !0 and as in Lemma 4.3, it follows that ~ 0 a0 max sign
e" f f^ ! 2
A:50
258
In addition, for any ^ 2 , there exists some 2 such that ~ 0 0 sign
e" f f^ !
implying that a0 0 and therefore a0
jmax j sign
e" f jj
~ 0 f^ !
Hence, from (A.50), we have that V_ 0 for all e" , thus proving the ~ ~ and ~. boundedness of ec , ,
References [1] Narendra, K. S. and Annaswamy. A. M. (1989). Stable Adaptive Systems. PrenticeHall, Inc., Englewood Clis, NJ. [2] KrsticÂ, M., Kanellakopoulos, I. and KokotovicÂ, P. V. (1994). `Nonlinear Design of Adaptive Controllers for Linear Systems', IEEE Transac. Autom. Contr., Vol. 39(4), 738±752. [3] Morse, A. S. (1993). `Supervisory Control of Families of Linear Set-point Controllers ± Part I: Exact Matching', Technical Report 9301, Yale University, March. [4] Armstrong-HeÂlouvry, B., Dupont, P. and Canudas de Wit, C. (1994). `A Survey of Models, Analysis Tools and Compensation Methods for the Control of Machines with Friction', Automatica, Vol. 30(7), 1083±1138. [5] Ting-Jen Yeh (1996). `Modeling, Analysis and Control of Magnetically Levitated Rotating Machines', PhD thesis, MIT, Massachusetts, MA. [6] BosÏ kovicÂ, J. D. (1994). `Stable Adaptive Control of a Class of Fed-batch Fermentation Processes`, in Proceedings of the Eighth Yale Workshop on Adaptive and Learning Systems, 263±268, Center for Systems Science, Dept. of Electrical Engineering, Yale University, New Haven, CT 06520-1968, June. [7] Fomin, V., Fradkov, A. and Yakubovich, V. (1981). Adaptive Control of Dynamical Systems. Eds. Nauka, Moscow. [8] Ortega, R. (1996). `Some Remarks on Adaptive Neuro-fuzzy Systems', International Journal of Adaptive Control and Signal Processing, Vol. 10, 79±83. [9] Annaswamy, A. M., Loh, A. P. and Skantze, F. P. (1998). `Adaptive Control of Continuous Time Systems with Convex/concave Parametrization', Automatica, Vol. 34, 33±49, January. [10] Loh, A. P., Annaswamy, A. M. and Skantze, F. P. (1997). `Adaptation in the Presence of a General Nonlinear Parametrization: An Error Model Approach, IEEE Transac. Autom. Contr. (submitted). [11] Annaswamy, A. M., Mehta, N., Thanomsat, C. and Loh, A. P. (1997). `Application of Adaptive Controllers Based on Nonlinear Parametrization to Chemical Reactors and Magnetic Bearings', in Proceedings of the ASME Dynamic Systems and Control Division, Dallas, TX, November .
259
[12] Skantze, F. P., Loh, A. P. and Annaswamy, A. M. (1997). `Adaptive Estimation of Discrete-time Systems with Nonlinear Parametrization', Automatica (submitted). [13] Armstrong-HeÂlouvry, B. (1991). Control of Machines with Friction, Kluwer Academic Publishers, Norwell, MA. [14] Hess, D. P. and Soom, A. (1990). `Friction at a Lubricated Line Contact Operating at Oscillating Sliding Velocities', J. of Tribology, Vol. 112(1), 147±152. [15] Canudas de Wit, C., Noel, P., Aubin, A. and Brogliato, B. (1991). `Adaptive Friction Compensation in Robot Manipulators: Low Velocities', International Journal of Robotics Research, Vol. 10(3), 189±199. [16] Viel, F., Jadot, F. and Bastin, G. (1997). `Robust Feedback Stabilization of Chemical Reactors', IEEE Transac. Autom. Contr., Vol. 42, 473±481. [17] Taylor, D. G., Kokotovic, P. V., Marino, R. and Kanellakopoulos, I. (1989). `Adaptive Regulation of Nonlinear Systems with Unmodeled Dynamics', IEEE Transac. Autom. Contr., Vol. 34, 405±412. [18] Kojic, A., Annaswamy, A M., Loh, A.-P. and Lozano, R. (1998). `Adaptive Control of a Class of Second Order Nonlinear Systems with Convex/concave Parameterization', in Conference on Decision and Control, Tampa, FL (to appear). [19] Seto, D., Annaswamy, A. M. and Baillieul, J. (1994). `Adaptive Control of Nonlinear Systems with a Triangular Structure', IEEE Transac. Autom. Contr., Vol. 39(7), 1411±1428, July. [20] Bakker, R. and Annaswamy, A. (1996). `Stability and Robustness Properties of a Simple Adaptive Controller', IEEE Transac. Autom. Contr., Vol. 41, 1352±1358.
10
Adaptive inverse for actuator compensation G. Tao
Abstract A general adaptive inverse approach is developed for control of plants with actuator imperfections caused by nonsmooth nonlinearities such as dead zone, backlash, hysteresis and other piecewise-linear characteristics. An adaptive inverse is employed for cancelling the eect of an actuator nonlinearity with unknown parameters, and a linear feedback control law is used for controlling the dynamics of a linear or smooth nonlinear part following the actuator nonlinearity. State feedback and output feedback control designs are presented which all lead to linearly parametrized error models suitable for developing adaptive laws to update the inverse parameters. This approach suggests that control systems with commonly used linear or nonlinear feedback controllers such as those with model reference, PID, pole placement or other dynamic compensation designs can be combined with an adaptive inverse for improving system tracking performance despite the presence of actuator imperfections.
10.1
Introduction
Adaptive control is a control methodology which provides adaptation mechanisms to adjust controllers for systems with parametric, structural and environmental uncertainties to achieve desired system performance. Payload variation or component ageing causes parametric uncertainties, component failure leads to structural uncertainties and external noises are typical environmental uncertainties. Such uncertainties often appear in control systems such as those in electrical, mechanical, chemical, aeronautical and biomedical engineering.
261
Adaptive control of linear systems has been extensively studied. Systematic design procedures have been developed for model reference adaptive control, pole placement control, self-tuning regulators and multivariable adaptive control. Robustness of adaptive control schemes with respect to modelling errors such as unmodelled dynamics, parameter variations and external disturbances has been a hot research topic. Recently adaptive controllers have been developed for nonlinear systems such as pure-feedback systems and feedback linearizable systems under the assumption that the nonlinearities are suciently smooth. Nonsmooth nonlinear characteristics such as dead zones, backlash and hysteresis are common in actuators, such as mechanical connections, hydraulic servo-valves, piezoelectric translators and electric servomotors, and also appear in biomedical systems [3, 5, 7, 8, 10, 18]. They are usually poorly known and may vary with time. They often severely limit system performance, giving rise to undesirable inaccuracy or oscillations or even leading to instability. The development of adaptive control schemes for systems with actuator imperfections has been a task of major practical interest. Recently, an adaptive inverse approach was proposed to deal with systems with nonsmooth actuator nonlinearities [10, 12]. The control scheme of [10] consists of an adaptive inverse for cancelling the eect of an unknown actuator nonlinearity such as a dead zone or a two-segment piecewise linearity and a linear state or output feedback design for a linear dynamics. In [12], the adaptive inverse approach is uni®ed for adaptive output feedback control of linear systems with unknown actuator or/and sensor dead zone, backlash and hysteresis, based on a model reference control method for systems with stable zeros. In this chapter, we describe how such an adaptive inverse approach can be combined with other control methods such as pole placement, PID, and how it may be applied to multivariable or nonlinear dynamics with actuator nonlinearities. In Section 10.2, we present a general parametrized actuator nonlinearity model illustrated by a dead-zone characteristic. In Section 10.3, we propose a parametrized inverse for cancelling the actuator nonlinearity, illustrated by a dead-zone inverse. In Section 10.4, we design state feedback adaptive inverse control schemes, while in Section 10.5, we develop a general output feedback adaptive inverse control scheme. In Section 10.6, we present three output feedback designs: model reference, pole placement, and PID (as illustrated by a backlash compensation example with simulation results), as examples of the general control scheme of Section 10.5. We also present feedback adaptive inverse control schemes in Section 10.7 for multivariable linear plants with actuator nonlinearities, using adaptive parameter update laws based on a coupled estimation error model or a Lyapunov design, and in Section 10.8 for smooth nonlinear dynamics with nonsmooth actuator nonlinearities, using an adaptive backstepping design.
262
10.2
Plants with actuator nonlinearities
Consider the plant with a nonlinearity N
, with unknown parameters, at the input of a known linear part G
D: y
t G
Du
t;
u
t N
v
t
10:1
where N
represents an actuator uncertainty such as a dead zone, backlash, hysteresis or piecewise-linear characteristic, v
t is the applied control, u
t is not accessible for either control or measurement and G
D is a rational transfer function either in continuous time (when D s denotes either the Laplace _ transform variable or the time dierentiation operator: sx
t x
t) or in discrete time (when D z denotes either the z-transform variable or the time advance operator: zx
t x
t 1), for a uni®ed presentation. The control objective is to design an adaptive compensator to cancel the eect of the uncertain actuator nonlinearity N
so that a commonly-used control scheme for the linear part G
D can be used to ensure desired system performance. To achieve such an objective, there are two key tasks: one is the clari®cation of the class of actuator nonlinearities for which such compensators can be developed, and the other is the design of adaptive laws which can eectively update the compensator parameters. In this section, we ful®l the ®rst task by presenting a parametrized nonlinearity model suitable for adaptive compensation schemes to be developed in the next sections. Dead-zone, backlash, hysteresis, and piecewise-linear characteristics are representatives of an actuator nonlinearity N
. These nonlinear characteristics have break points so that they are nondierentiable (nonsmooth) but they can be parametrized [10, 12, 15]. The parametrized models of these nonlinearities can be uni®ed as u
t N
v
t N
; v
t
T !
t a
t
10:2
where 2 Rn (n 1) is an unknown parameter vector, and !
t 2 Rn and a
t 2 R whose components are determined by the signal motion in the nonlinear characteristic N
and therefore are also unknown. To illustrate the nonlinearity model (10.2), let us consider a dead-zone characteristic DZ
with the input-output relationship: > < mr
v
t br if v
t br u
t N
v
t DZ
v
t 0
10:3 if bl < v
t < br > : ml
v
t bl if v
t bl where mr > 0, ml > 0, br > 0, and bl < 0 are dead zone parameters.
Introducing the indicator function X of the event X: 1 if X is true X 0 otherwise
263
10:4
we de®ne the dead-zone indicator functions r
t u
t > 0
10:5
l
t u
t < 0
10:6
Then, introducing the dead-zone parameter vector and its regressor
mr ; mr br ; ml ; ml bl T !
t
r
tv
t; r
t; l
tv
t; l
t
10:7 T
10:8
we obtain (10.2) with a
t 0 for the dead-zone characteristic (10.3). For a parametrized nonlinearity N
in (10.2), we will develop an adaptive inverse as a compensator for cancelling N
with unknown parameters.
10.3 Parametrized inverses The essence of the adaptive inverse approach is to employ an inverse c d
t v
t NI
u
10:9
to cancel the eect of the unknown nonlinearity N
, where the inverse c characteristic NI
is parametrized by an estimate of , and ud
t is a desired control signal from a feedback law. The key requirement for such an inverse is that its parameters can be updated from an adaptive law and should stay in a prespeci®ed region needed for implementation of an inverse. In our designs, such an adaptive law is developed based on a linearly parametrized error model and the parameter boundaries are ensured by parameter projection.
A desirable inverse (10.9) should be parametrizable as ud
t
T
t!
t a
t
10:10
n
for some known signals !
t 2 R and a
t 2 R whose components are c such that determined by the signal motion in the nonlinearity inverse NI
v
t, !
t and a
t are bounded if ud
t is. The error signal (due to an uncertain N
) dn
t T
!
t !
t a
t a
t
10:11 should satisfy the conditions that dn
t is bounded, t 0, and that dn
t 0, c is correctly initialized: dn
t0 0. t t0 , if
t , t t0 , and NI
264
As shown in [10, 12, 15], the inverses for a dead zone, backlash, hysteresis and piecewise linearity have such desired properties. Here we use the dead-zone inverse as an illustrative example. Let the estimates of d d cr , m cl , respectively. Then the inverse for the mr br , mr , ml bl , ml be m r br , m l bl ; m dead-zone characteristic (10.3) is described by 8 d u
t m > r br > > d if ud
t > 0 > > c m r < c d
t DI
u c d
t
10:12 vt NI
u 0 if ud
t 0 > > > u
t m d > > l bl : d if ud
t < 0 cl m For the dead zone inverse (10.12), to arrive at the desired form (10.10), we introduce the inverse indicator functions br
t v
t > 0
10:13
bl
t v
t < 0
10:14
and the inverse parameter vector and regressor T d d cl ; m
c mr ; m r br ; m l bl
10:15 T
!
t
br
tv
t; br
t; bl
tv
t; bl
t
10:16
Then, the dead-zone inverse (12) is cr br
tv
t ud
t m
T !
t
d cl bl
tv
t br
t m m r br
d bl
t m l bl
10:17
that is, a
t 0 in (10.10). It follows from (10.7), (10.8), (10.11) and (10.16) that dn
t T !
tu
t 0
mr 0 < v
t < br
v
t
br
ml bl < v
t < 0
v
t
bl
10:18
which has the desired properties that dn
t is bounded for all t 0 and that dn
t 0 whenever . Furthermore, dn
t 0 whenever v
t br or v
t bl , that is, when u
t and v
t are outside the dead zone, which is the d d m m r br l bl case when bbr X br and bbl X bl , see [12]. cr cl m m
It is important to see that the inverse (10.10), when applied to the nonlinearity (10.2), results in the control error u
t
ud
t
T !
t dn
t
10:19
265
which is suitable for developing an adaptive inverse compensator. For the adaptive designs to be presented in the next sections, we assume that the inverse block (10.9) has the form (10.10) and dn t in (11) has the stated properties. We should note that the signals a
t in (10.2) and a
t in (10.10) are non zero if the nonlinearity N
and its inverse NI
have inner loops as in the hysteresis case [12].
10.4 State feedback designs Consider the plant (10.1), where G
D has a controllable state variable realization Dx
t Ax
t Bu
t
10:20 y
t Cx
t for some known constant matrices A 2 Rnn , B 2 Rn1 and C 2 R1n , n > 0, that is, G
D C
DI A 1 B, and r
t is a bounded reference input signal. In the presence of an actuator nonlinearity u
t N
v
t parametrizable as in (10.2), we propose the control scheme shown in Figure 10.1, which consists c d
t parametrizable as in (10.10) and the state of the inverse v
t NI
u feedback law ud
t Kx
t r
t
10:21 where K 2 R1n is a constant gain vector such that the eigenvalues of A BK are equal to some desired closed loop system poles. The choice of such a K can be made from a pole placement or linear quadratic optimal control design. A pole placement design can be used to match the closed loop transfer function C
DI A BK 1 B to a reference model if the zeros of G
D have good stability.
With the controller (10.21), the closed loop system is
y
t Wm
Dr
t Wm
D
T ! dn
t
t
10:22
where
t is an exponentially decaying term due to initial conditions, and Wm
D C
DI
A
BK 1 B
10:23
The ideal system output is that of the closed loop system without the actuator c nonlinearity N
and its inverse NI
, that is, with u
t ud
t. In this case y
t Wm
Dr
t
t:
10:24
In view of this output, we introduce the reference output ym
t Wm
Dr
t:
10:25
Ignoring the eect of
t, introducing d
t Wm
Ddn
t which is bounded because dn
t in (10.11) is, and using (10.22) and (10.25), we have
266
r
ud
c NI
v
N
u
Dx Ax Bu
x
C
y
K
Figure 10.1
yt ym t Wm
D
T !
t d
t
Wm
DT !
t
T Wm
D!
t d
t
10:26
In view of (10.26), we de®ne the estimation error "
t y
t
ym
t
t
10:27
where
t T
t
t
Wm
DT !
t
t Wm
D!
t
10:28
10:29
Using (10.26)±(10.29), we arrive at the error equation "
t
t
T
t d
t
10:30
which is linear in the parameter error
t , with an additional bounded error d
t due to the uncertainty of the nonlinearity N
in the plant (10.1). The adaptive update law for
t
1
t; . . . ; n
tT can be developed using a gradient projection algorithm to minimize a cost function J
. The related optimization problem is to ®nd an iterative algorithm to generate a sequence of which will asymptotically minimize J
; s.t. i 2 ai ; bi ;
i 1; 2; . . . ; n
10:31
for some known parameter bounds ai < bi such that i 2 ai ; bi , i 1; 2; . . . ; n , for
i ; . . . ; n T , and that for any set of i 2 ai ; bi , i 1; 2; . . . ; p, the characteristic N
; renders an actuator uncertainty model in the same class to which the true model N
; belongs, for example, for N
; being a dead zone of the form (10.3), N
; should also be a deadzone of the form (10.3) with possibly dierent parameters. The constraint that c is i 2 ai ; bi , i 1; 2; . . . ; n ; is to ensure that the nonlinearity inverse NI
implementable as an inverse for the type of nonlinearity N
under consideration.
267
In the dead-zone case, the parameter boundaries are mr1 mr mr2 , ml1 ml ml2 , 0 br br0 , bl0 bl 0, for some known positive constants mr1 , ml1 , mr2 , ml2 , br0 , bl0 . The above constraint condition, in terms of
mr ; mr br ; ml ; ml bl T as in (10.7), is i 2 ai ; bi ; i 1; 2; 3; 4, for a1 mr1 ; a2 0; a3 ml1 ; a4 ml2 bl0 , b1 mr2 ; b2 mr2 br0 ; b3 ml2 ; b4 0: Then for any i 2 ai ; bi ; i 1; 2; 3; 4, DZ
; is a dead-zone characteristic. For the formulated optimization problem, the gradient projection method [4, 9, 12] suggests the adaptive update law for
t: _ @J
t
t f
t
10:32 @
t 1
t where f
t is a parameter projection term, and diag f 1 ; ; n g
10:33
with i > 0 for a continuous-time design and 0 < i < 2for a discrete-time design. To ensure i 2 ai ; bi , i 1; 2; . . . ; n , we introduce g
t
g1
t; . . . ; gn
tT
@J
t @
10:34
initialize the adaptive law (10.35) with i
0 2 ai ; bi ; i 1; 2; . . . ; n and set the ith component of the 8 0 > > > < fi
t > > > : gi
t
10:35
modi®cation term f
t as if i
t 2
ai ; bi if i
t ai ; gi
t 0; or if i
t bi ; gi
t 0
10:36
otherwise
in a continuous-time design, or 8 if i
t gi
t 2 ai ; bi > bi > : a i i
t gi
t if i
t gi
t < ai
10:37
in a discrete-time design, for i 1; 2; . . . ; n . For the commonly used cost function J
where "
"2 2m2
T d (see (10.30)), and p m
t 1 T
t
t
10:38
10:39
268
such that
2 is bounded if t is bounded, it follows that 2m2 @J "
t
t
t 2 @ m
t
Therefore, a desired adaptive update law for
t is ) _ "
t
t
t f
t m2
t
t 1
t
10:40
10:41
initialized by (10.35) and projected with f
t in (10.36) or (10.37) for "
t
t g
t . m2
t "
t This adaptive law ensures that i
t 2 ai ; bi , i 1; 2; . . . ; n , and that m
t _ or
t 1
t are bounded by d
t in a mean square sense. and
t m
t ~
t , In continuous time, with
t the closed loop system (10.20)±(10.21) is ~ _
A BKx
t Br
t B
t!
t x
t dn
t y
t Cx
t
10:42
The ideal system performance is that with u
t ud
t which means ~
t!
t dn
t 0. This motivates us to introduce the reference system x_ m
t
A BKxm
t Br
t ym
t Cxm
t
10:43
De®ning the state error vector x~
t x
t xm
t and the output error e
t y
t ym
t, from (10.42) and (10.43), we obtain ~ x~_
t
A BK~ x
t B
t!
t dn
t e
t Cx~
t
10:44
Consider the positive de®nite function T ~ 1
~ V
~ x; x ~T 2 x P~
1~
10:45
where P 2 Rnn with P PT > 0 satisfying the Lyapunov equation P
A BK
A BKT P for some n n matrix Q QT > 0, and
2Q
10:46
diag f 1 ; ; n g with i > 0,
269
~ is i 1; 2; . . . ; n . With (10.44) and (10.46), the time derivative of V
~ x; d ~ _ x; V
t V
~ dt x~T
tPx~_
t ~T
t
1 _~
t
_~ x~T
tQ~ x
t x~T
tPB
~T
t!
t dn
t ~T
t 1
t
10:47 ~_
t _ !
t~ xT
tPB would If dn
t were absent, then the choice of
t T _ x
t 0, as desired. In the presence of dn
t and with the make V
t x~
tQ~ need of parameter projection, we choose the adaptive update law for
t as
_
t
!
t~ xT
tPB f
t
10:48
initialized by (10.35) and projected with f
t in (10.36) for g
t !
t~ xT
tPB. With the adaptive law (10.48), we have _ V
t
x~T
tQ~ x
t 2~ xT
tPBdn
t 2~T
t
1
f
t
10:49
From the parameter projection algorithm (10.36), it follows that i
t 2 ai ; bi ;
i
t
i fi
t
0;
i 1; 2; . . . ; n
i 1; 2; . . . ; n
10:50
10:51
that is, ~T
t 1 f
t 0. Since Q QT > 0 and dn
t is bounded, we have, from (10.49) and (10.51), the boundedness of x~
t, from (10.43), that of xm
t, and in turn, from x~
t x
t xm
t, that of x
t, from (10.21), that of ud
t, and from (10), that of !
t, a
t and v
t. Thus all closed-loop signals are bounded. Finally, from (10.48), (10.49) and (10.51), it can be veri®ed that x~
t, _ are all bounded by dn
t in a mean square sense. e
t y
t ym
t and
t
10.5 Output feedback inverse control When only the output y
t of the plant (10.1) is available for measurement, we need to use an output feedback control law to generate ud
t for the inverse (10.9). As shown in Figure 10.2, such a controller structure is ud
t C1
Dw
t; w
t r
t
C2
Dy
t
10:52
where r
t is a reference input, and C1
D and C2
D are linear blocks which are designed as if the nonlinearity N
were absent, that is, when u
t ud
t. The desired reference output ym
t to be tracked by the plant output y
t now is de®ned as the output of the closed loop system with the controller (10.21) applied to the plant (10.1) without the actuator non-
270
linearity N
, that is, when u
t ud
t. With u
t ud
t, the closed loop output is y
t ym
tX
G
DC1
D r
t 1 G
DC1
DC2
D
10:53
Hence C1
D and C2
D should be chosen such that the transfer function Wm
D
G
DC1
D 1 G
DC1
DC2
D
10:54
has all its poles at some desired stable locations, and that the closed loop system is internally stable when u
t ud
t. To develop an adaptive law for updating the parameters of the c inverse NI
, we express the closed loop system as y
t ym
t W
D
where
W
D
T !
t d
t
G
D 1 G
DC1
DC2
D
10:55
10:56
d
t W
Ddn
t
10:57
To use (10.55) to derive a desirable error model, W
D in (10.56) is required to have all its poles located in a desired stable region of the complex plane. With this requirement, the error signal d
t in (10.57) is bounded because dn
t in (10.11) is. Similar to (10.26)±(10.28), in view of (10.55), we de®ne the estimation error "
t y
t where
ym
t
t
t T
t
t
10:58
W
DT !
t
10:59
t W
D!
t ud
w r
C1
D
c NI
10:60
v
C2
D
Figure 10.2
N
u
G
D
y
271
Using (10.55), (10.58)±(10.60), we arrive at the error equation t
T
t d
t
10:61
which has the same form as that in (10.30) for a state feedback design. Based on the linearly parametrized error equation (10.61) and the gradient projection optimization method described in Section 10.4, similar to that in (10.41), we choose the adaptive update law for
t: ) _
t
t
t f
t
10:62 m2
t
t 1
t which is initialized by (10.35) and projected with f
t in (10.36) or (10.37) for
t
t , where is given in (10.33). This adaptive law also ensures g
t m2
t "
t and that i
t 2 ai ; bi , i 1; 2; . . . ; n , and that the normalized error m
t d
t _ in a mean parameter variation
t or
t 1
t are bounded by m
t sense.
An output feedback adaptive inverse control design for the plant (10.1) with an actuator nonlinearity N
requires: (i) N
should be parametrized by
1 ; . . . ; n T as in (10.2); c in (10.9), as a compensator to cancel N
, (ii) the nonlinearity inverse NI
should be parametrizable as (10.10) with the stated properties; (iii) the true nonlinearity parameters i 2 ai ; bi for some known constants ai < bi , i 1; 2; . . . ; n , such that for any set of i 2 ai ; bi , i 1; 2; . . . ; n , the characteristic N
; renders a nonlinearity model of the same type as that of the true nonlinearity model N
; ; and (iv) the compensators C1
D and C2
D in the controller (10.52) should ensure that both Wm
D in (10.54) and W
D in (10.56) have good stability properties. As a comparison, the conditions (i)±(iii) are also needed for a state feedback adaptive inverse control design in Section 10.4, while the state feedback control law (10.21) can always place the eigenvalues of A BK at some desired locations under the controllability condition on the system matrices
A; B in (10.20).
10.6 Output feedback designs Many existing control design methods for the plant (10.1) without the actuator
272
nonlinearity N
may be used to construct the compensators C1
D and C2
D in the controller (10.52) to be incorporated in the adaptive inverse control scheme developed in Section 10.5 for a plant with an actuator nonlinearity. Those designs include model reference, PID and pole placement which will be given in this section. To parametrize the linear output feedback controller Z
D , where kp is a constant gain, and Z
D and (10.52), we express G
D kp P
D P
D are monic polynomials of degrees n and m, respectively. Using the notation D, we present continuous- and discrete-time designs in a uni®ed framework. !
[12]. A model reference control design for (10.52) is
T ud
t T 1 !1
t 2 !2
t 20 y
t 3 r
t
10:63
where !1
t, !2
t are the signals from two identical ®lters: !1
t
a
D a
D ud
t; !2
t y
t
D
D
10:64
with a
D
1; D; . . . ; Dn 2 T and
D being a stable polynomial of degree n 1, and 1 2 Rn 1 ; 2 2 Rn 1 ; 20 2 R; 3 2 R satisfy 3 kp 1 and T T 1 a
DP
D
2 a
D 20
Dkp Z
D
D
P
D
Z
DPm
D
10:65
where Pm
D is a stable polynomial of degree n m. In the form of the controller (10.52), we identify 3 T a
D 20 ; C2
D kp 2 C1
D a
D
D 1 T 1
D With this choice of C1
D and C2
D, we have kp 1 ; W
D Wm
D 1 Pm
D Pm
D
T 1
a
D
D
10:66
10:67
It shows that both Wm
D and W
D have good stability properties from that of Pm
D and
D independent of G
D. For internal stability, the zero polynomial Z
D should be stable because model reference control cancels the dynamics of Z
D to lead to the closed loop system y
t Wm
Dr
t in the c absence of the nonlinearity N
and its inverse NI
, that is, when u
t ud
t.
"
[14]. A pole placement control design is similar:
T ud
t T 1 !1
t 2 !2
t 20 y
t 3 r
t
10:68
273
where !1 t and !2
t are the same as that de®ned in (10.64), and 1 2 Rn 1 ; 2 2 Rn 1 ; 20 2 R; 3 2 R satisfy 3 kp 1 and T T 1 a
DP
D
2 a
D 20
Dkp Z
D
D
P
D
Pd
D
10:69
where Pd
D is a stable polynomial of degree n which contains the desired closed loop poles. In this design, C1
D and C2
D are also expressed in (66) but kp Z
D Z
D a
D Wm
D ; W
D
10:70 1 T 1 Pd
D
D Pd
D It also shows that both Wm
D and W
D have good stability properties from that of Pd
D and
D independent of G
D. However, the reference output ym
t Wm
Dr
t now depends on the plant zero polynomial Z
D which is allowed to be unstable. To solve (10.69) for an arbitrary but stable Pd
D, it is required that Z
D and P
D are co-prime. "#
A typical PID design (10.52) has C1
D C
D and C2
D 1: ud
t C
Dr
y
t
10:71
s s
10:72
where the PID controller C
D has the form C
D C
s for a continuous-time design, and C
D C
z
z
1
z
1 z
10:73
for a discrete-time design, where , and are design parameters. Unlike either a model reference design or a pole placement design, which can be systematically done for a general nth-order plant G
D, a PID controller only applies to certain classes of plants. In other words, given a linear plant G
D, one needs to make sure that a choice of C
D is able to ensure both Wm
D in (10.54) and W
D in (10.56) with good stability properties. Many simulation results for model reference and pole placement designs were presented in [12, 14, 15], which indicate that control systems with an adaptive inverse have signi®cantly improved tracking performance, in addition to closed-loop signal boundedness. As an additional illustrative example, we now present a detailed PI design. An example of systems with actuator backlash is a liquid tank where the backlash is in the valve control mechanism. The liquid level y
t is described as y
t G
su
t; u
t B
v
t
10:74
274
k with k being a constant (k 1 for s simplicity), B
represents the actuator backlash and v
t is the control. A simple version of the backlash characteristic u
t B
v
t is described by two parallel straight lines connected with inner horizontal line segments of length 2c > 0. The upward line is active when both v
t and u
t increase:
where ut is the controlled ¯ow, G
s
u
t v
t
c;
v_
t > 0;
10:75
_ >0 u
t
the downward line is active when both v
t and u
t decrease: _ 0
10:79
Vertical jumps of v
t occur whenever u_ d
t changes its sign. Introducing the backlash parameter and its estimate c;
t cb
t
10:80
and de®ning the backlash inverse regressor 1 if v
t ud
t cb
t b
t !
t 2b
t 1; 0 otherwise
10:81
we have the control error expression (10.19), where dn
t 2 2c; 2c and dn
t 0, t t0 , if and the backlash inverse is correctly initialized: c has the desired u
t0 ud
t0 . In other words, the backlash inverse BI
property: c ; ud
t0 ud
t0 B
BI
)
c ; ud
t ud
t; B
BI
8t t0
10:82
Combined with the backlash inverse (10.77) is the PI controller r y
t; > 0; > 0 ud
t s
10:83
275
which leads to the closed loop error system yt
s s r
t 2
s2 s s s
! dn
t
10:84
s s
10:85
In view of (10.53), (10.55) and (10.84), we have Wm
s
s2
s ; s
W
s
s2
which have desired stability properties for some and . Then, we can update
t from the continuous-time adaptive law (10.62) with f
t chosen to ensure b c
t 2 c1 ; c2 where 0 < c1 < c < c2 with c1 and c2 being known constants. For the same choice of 1, r 10, and y
0 5 as those led to large tracking error e
t without a backlash inverse [12], what happens in the closedc loop system with an adaptive backlash inverse BI
is shown in c
0 0:5 (other choices of cb
0 led to similar results). Figure 10.3(a)±(f) for b These results clearly show that an adaptive backlash inverse is able to adaptively cancel the eects of the unknown backlash so that the system tracking performance is signi®cantly improved: the tracking error e
t y
t r reduces to very small values, and the limit cycle, which appeared with a PI controller alone [12], is eliminated.
10.7 Designs for multivariable systems Many control systems are multivariable. Coupling dynamics in a multivariable system often make it more dicult to control than a single-variable system. In this section, state feedback and output feedback adaptive inverse control schemes are presented for compensation of unknown nonsmooth nonlinearities at the input of a known multivariable linear plant, based on a multivariable parameter estimation algorithm for coupled error models. Consider a multivariable plant with actuator nonlinearities, described by y
t G
Du
t;
u
t N
v
t
10:86
where G
D fgij
Dg is an m m strictly proper transfer matrix, u
t
u1
t; . . . ; um
tT and v
t
v1
t; . . . ; vm
tT are the output and input of the multivariable actuator nonlinearity N
N1
; . . . ; Nm
T such that
10:87 ui
t Ni
vi
t; i 1; 2; . . . ; m The control problem is to design adaptive inverses for cancelling the nonlinearities Ni
, to be combined with commonly used multivariable control schemes for the linear part G
D for desired tracking performance. In this problem, v
t is the control and u
t is not accessible for either control or measurement.
276
(a) e
t vs. V
t
(b) ut vs. v
t
(c) Tracking error et
(d) Parameter error c^
t c
(e) Backlash output u
t
(f) Backlash input v
t
Figure 10.3
c^
0 0:5
277
Similar to that in Section 10.2 for m 1, we assume that each Ni
, which may be a dead-zone, backlash, hysteresis or other piecewise-linear characteristic, can be parametrized as ui
t Ni
vi
t Ni
i ; vi
t
T i !i
t ai
t
10:88
for some unknown parameter vectors i 2 Rni , ni 1, i 1; . . . ; m, and some unknown regressor vector signals !i
t 2 Rni and scalar signals ai
t.To cancel such nonlinearities, we use a multivariable nonlinearity inverse c d
t v
t NI
u
10:89
where ud
t
ud1
t; . . . ; udm
tT is a design vector signal to be generated for c
NI c1
; . . . ; NI cm
T , that is, the inverse (10.89) is equivalent to NI
ci
udi
t; vi
t NI
i 1; . . . ; m
10:90
ci
can be parametrized as We also assume that each NI udi
t
Ti
t!i
t ai
t;
i 1; . . . ; m
10:91
where i 2 Rni is an estimate of i , and !i
t 2 Rni and ai
t are some known signals, as in the case of an inverse for a dead zone, backlash or hysteresis. The uncertainties in Ni
causes a control error at each input channel: ui
t where ~i
t i
t
udi
t ~Ti
t!i
t di
t
10:92
i , i 1; 2; . . . ; m, and the unparametrized error is di
t T i
!i
t
!i
t ai
t
ai
t
10:93
which should satisfy the same condition as that for dn
t in (10.11). In the vector form, the control error (10.92) is u
t
~ T
t!
t dn
t ud
t
10:94
where 0
~T1
t
B B ~ T
t B 0 B @ 0 0
0
0
~T2
t
0
0
0
~Tm
t
!
t
!T1
t; . . . ; !Tm
tT
dn
t
d1
t; . . . ; dm
t
T
1 C C C C A
10:95
10:96
10:97
278
$%
Let the plant (10.86) be in the state variable form: Dx
t Ax
t Bu
t y
t Cx
t
10:98
where A 2 Rnn , B 2 Rnm and C 2 Rmn , with
A; B being controllable. To generate ud
t for the the inverse (10.89), we use the state feedback controller ud
t Kx
t r
t
10:99
where K 2 Rmn is such that the eigenvalues of A BK are placed at some desired locations, and r
t is a bounded reference input signal. Using (10.94), (10.98) and (10.99), we obtain T
t!
t Bdn
t Dx
t
A BKx
t Br
t B y
t Cx
t
10:100
~ T
t!
t dn
t 0. The ideal system performance is that of (10.100) with Hence, we can de®ne the reference output ym
t for y
t as ym
t Wm
Dr
t
10:101
where Wm
D C
DI
A
BK 1 B
10:102
Then, with d
t Wm
Ddn
t, the output tracking error is e
t
e1
t; . . . ; em
tT y
t
~ T !
t d
t ym
t Wm
D
10:103
Since Wm
D is an m m transfer matrix, the error equation (10.103) has coupled dynamics, for which the parameter estimation algorithm of [13] can be applied. To proceed, we de®ne and the estimation errors "i
t ei
t i1
t . . . im
t
10:104
where ij
t Tj
tij
t
wij
DTj !j
t
ij
t wij
D!j
t
10:105
10:106
with wij
D being the
i; jth element of Wm
D, for i; j 1; 2; . . . ; m. Substituting (10.103), (10.105), (10.106) in (10.104) gives "i
t
m X
j
t j1
j T ij
t di
t
10:107
279
where di is the ith component of dt. Introducing
t
T1
t; . . . ; Tm
tT
i
t
10:108
T T
T 1 ; . . . ; m
10:109
T T
i1
t; . . . ; im
tT
10:110
we express the error model (10.107) as "i
t
t
T i
t di
t;
i 1; 2; . . . ; m
10:111
This is a set of linear error models with bounded `disturbances' di
t. One important feature of these equations is that dynamic coupling in the multivariable error equation (10.103) leads to a set of m estimation errors which all contain the overall parameter vectors
t and . Introducing the cost function "21 . . . "2m 2m2 m X m2
t 1 iT
ti
t
10:112
J
10:113
i1
and applying the gradient projection optimization technique of Section 10.4, we have the adaptive law for
t: m X _
t "i
ti
t f
t
10:114 m2
t i1
t 1
t where f
t is for parameter projection and is an adaptation gain matrix. Both f
t and can be chosen similar to that in (10.36) or (10.37) and (10.33). To develop the multivariable counterpart of the continuous-time adaptive scheme (10.48), with the same notation as that in (10.42)±(10.46) except that ~ T
t!
t, we have the time derivative of V
~ ~ in ~T
t!
t is replaced by x; (10.45) as _ V
t
~ T
t!
t dn
t ~T
t x~T
tQ~ x
t x~T
tPB
~
t where
t
1 _~
t
10:115
in (10.108) and (10.109). With the known signals xT PBT x
t
x1
t; . . . ; xm
tT
~
10:116
!
t x1
t!T1
t; . . . ; xm
t!Tm
tT
10:117
we express (10.115) as _ V
t
x~T
tPBdn
t ~T
t x
t ~T
t!
t x~T
tQ~
1 _~
t
10:118
280
which suggests the following adaptive law for t: _
t
f
t !
t
where f
t is for parameter projection and
10:119
is diagonal and positive.
& $% To develop a multivariable output feedback adaptive inverse control scheme, we assume that G
D is nonsingular with the observability index v and has the form: G
D Z
DP 1
D, for some m m known polynomial matrices Z
D and P
D. The controller structure for ud
t is T ud
t T 1 !1
t 2 !2
t 20 y
t 3 r
t
where !1
t
A
D ud
t; n
D
!2
t
10:120
A
D y
t n
D
A
D
I; DI; . . . ; D 2 IT
10:121
10:122
n
D is a monic and stable polynomial of degree 2 2 Rm
1m , 20 2 Rmm , and 3 2 Rmm satisfy
1, and 1 2 Rm
T T 1 A
DP
D
2 A
D 20 n
DA
DN
D n
D
P
D
1m
,
3 Pd
D
10:123
for a given m m stable polynomial matrix Pd
D. A model reference design needs a stable Z
D. The choice to meet (10.123) is: Pd
D m
DZ
D and 3 Kp 1 , where m
D is a modi®ed interactor matrix (which is a stable polynomial matrix) of G
D such that lim m
DG
D Kp
10:124
D!1
is ®nite and nonsingular [11]. For a pole placement design which does not need a stable Z
D, 3 Pd
D is chosen to have a structure similar to that of P
D so that (10.123) has a solution under the right co-primeness of Z
D and P
D [1]. With the controller (10.126) and the inverse (10.89), we have the error equation ~ T !
t d
t e
t y
t ym
t W
D
10:125 where
ym
t Z
DPd 1
Dr
t W
D Z
DPd 1
D3 1 I
10:126 T 1
A
D n
D
10:127
and d
t W
Ddn
t is bounded because dn
t is and W
D is stable. Since the error equation (10.125) has the same form as that in (10.89), similar to (10.104)±(10.114), one can also develop an adaptive update law for i
t, i 1; . . . ; m.
281
10.8 Designs for nonlinear dynamics The adaptive inverse approach developed in Sections 10.4±10.6 can also be applied to systems with nonsmooth nonlinearities at the inputs of smooth nonlinear dynamics [16, 17]. State feedback and output feedback adaptive inverse control schemes may be designed for some special cases of a nonlinear plant _ f
x
t g
x
tu
t; u
t N
v
t xt y
t h
x
t
10:128
where f
x 2 Rn , g
x 2 Rn , and h
x 2 R are smooth functions of x 2 Rn , N
represents an unknown nonsmooth actuator uncertainty as a dead-zone, backlash, hysteresis or piecewise-linear characteristic, and v
t 2 R is the control input, while u
t is not accessible for either control or measurement. The main idea for the control of such plants is to use an adaptive inverse c d
t v
t NI
u
10:129
to cancel the eect of the unknown nonlinearity N
so that feedback control c schemes designed for (10.128) without N
, with the help of NI
, can be applied to (10.128) with N
, to achieve desired system performance. To illustrate this idea, we present an adaptive inverse design [16] for a thirdorder parametric-strict-feedback nonlinear plant [6] with an actuator nonlinearity N
: x_ 1 x2 T s '1
x1 x_ 2 x3 T s '2
x1 ; x2 x_ 3 '0
x
T s '3
x
10:130
0
xu;
u N
v
where x
x1 ; x2 ; x3 T , xi , i 1; 2; 3, are the state variables available for measurement, s 2 Rns is a vector of unknown constant parameters with a known upper bound Ms : ks k2 < Ms , for the Euclidean vector norm k k2 , '0 2 R, 0 2 R and i 2 Rp , i 1; 2; 3, are known smooth nonlinear functions, and b1 < j0
xj < b2 , for some positive constants b1 , b2 and 8x 2 R3 . To develop an adaptive inverse controller for (10.130), we assume that the c nonlinearity N
is parametrized by 2 Rn as in (10.2) and its inverse NI
is parametrized by 2 Rn , an estimate of , as in (10.10), with the stated properties. Then, the adaptive backstepping method [6] can be combined with c to control the plant (10.130), in a three-step design: an adaptive inverse NI
Step 1: Let the desired output be r
t to be tracked by the plant output y
t, with bounded derivatives r
k
t, k 1; 2; 3. De®ning z1 x1 r and z2 x2 1 , where 1 is a design signal to be determined, we have z_ 1 z2 1 T s '1
r_
10:131
282
to be stabilized by 1 with respect to the partial Lyapunov function s T
V1 12 z21 12
s
s
1
s
s
10:132
where s
t 2 Rns is an estimate of the unknown parameter vector s , and T s s > 0. The time derivative of V1 is V_ 1 z1
z2 1 Ts '1
_
s r
s T
1 s
_s
s z1 ' 1
10:133
Choosing the ®rst stabilizing function as 1 with c1 > 0, and de®ning 1 V_ 1 Step 2:
Ts '1 r_
c1 z1
s z1 ' 1 ,
we obtain
c1 z21 z1 z2
s
Introducing z3 x3
z_2 z3 2 T s '2
10:134
s T
1 s
_s
1
10:135
2 with 2 to be determined, we have
@1
x2 T s '1 @x1
@1 _ s @s
@1 r_ @r
@1 r @ r_
10:136
To derive 2 , together with 1 in (10.134), to stabilize (10.131) and (10.136) with respect to V2 V1 12 z22 , we express the time derivative of V2 as @1 @1 _ @1 @1 r x2 V_ 2 c1 z21 z2 z1 z3 2 r_ s @x1 @s @r @ r_ @1 T '1
s s T s 1 _s 2
10:137 s
'2 @x1 @1 '1 . Suggested by (10.137), we choose the where 2 1 s z2 '2 @x1 second stabilizing function 2 as 2
@1 @1 x2
2 s fs
s s @x1 @s @1 @1 @1 T r s '2 '1 r_ @r @ r_ 1 z1
c2 z 2
10:138
where c2 > 0, and fs
s
(
if ks k2 Ms
0 f0
1
e
0
ks k2 Ms 2
if ks k2 > Ms
10:139
with f0 > 0, 0 > 0, and Ms > ks k2 . Then we can rewrite (10.137) as
V_ 2
c1 z21
s
s T
1 s
z2
_s
@1 _ s @s
2
s fs
s s
2
10:140
Choose the feedback control ud
t for (10.129) as
Step 3: ud
c2 z22 z2 z3
283
1 0
'0
c3 z3
z2
2 X @2 k1
Ts
z2
'3 s
@1 @s
@xk
xk1
2 X @2 k1
@xk
'k
3 X @2 _ @2
k r s @s @r
k 1 k1 s fs
s s z2
@1 @s
10:141
where c3 > 0, and the adaptive update law for s
t is _s
s
3 X
s fs
s s
z l !l
10:142
l1
with !i
x1 ; . . . ; xi ; s 'i
i 1 X @i
1
@xk k1 With this choice of ud , we have z_3
z2
c3 z3
'k , for i 1; 2; 3 and 0 0.
s
@1 @s
s T !3 z2
s !3
10:143
Then, the overall system Lyapunov function V V2 12 z23 12
with V_
T
1
10:144
being diagonal and positive, has the time derivative 3 X
ck z2k
fs
s
s
s T s
T
1
_ z3 0 ! z3 0 dn
k1
10:145 n
where !
t 2 R is the known vector signal describing the signal motion of the c in (10.10), and dn
t is the bounded unparametrized nonlinearity inverse NI
error in (10.11). The expression (10.145), with the need of parameter projection c for
t for implementing an inverse NI
, suggests the adaptive update law for
t:
10:146 _ z3 0 ! f initialized by (10.35) and projected with f
t in (10.36) for g
t z3 0 !. This adaptive inverse control scheme consisting of (10.129), (10.141), (10.142) and (10.146) has some desired properties. First, the modi®cation (10.139) ensures that fs
s
s s T s 0, and the parameter projection (10.36) ensure that
T 1 f 0 and i
t 2 ai ; bi , i 1; . . . ; n , for
284
t
1
t; . . . ; n
tT and ai , bi de®ned in (10.31). Then, from (10.145) and (10.146), we have V_
3 X
ck z2k
fs
s
s
s T s
T
1
f z 3 0 dn
10:147
k1
Since dn and 0 are bounded, and fs
s
s s T s grows unboundedly if s grows unboundedly, it follows from (10.144) and (10.147) that s ; 2 L1 , and zk 2 L1 , k 1; 2; 3. Since z1 x1 r is bounded, we have x1 2 L1 and hence '1
x1 2 L1 . It follows from (10.134) that 1 2 L1 . By z2 x2 1 , we have x2 2 L1 and hence '2
x1 ; x2 2 L1 . It follows from (10.138) that 2 2 L1 . Similarly, x3 2 L1 , and ud 2 L1 in (10.141). Finally, v
t in (10.129) and u
t in (10.128) are bounded. Therefore, all closed loop system signals are bounded. Similarly, an output feedback adaptive inverse control scheme can be developed for the nonlinear plant (10.128) in an output-feedback canonical form with actuator nonlinearities [17]. For such a control scheme, a state observer [6] is needed to obtain a state estimate for implementing a feedback control law to generate ud
t as the input to an adaptive inverse: c d
t, to cancel an actuator nonlinearity: u
t N
v
t. v
t NI
u
10.9
Concluding remarks
Thus far, we have presented a general adaptive inverse approach for control of plants with unknown nonsmooth actuator nonlinearities such as dead zone, backlash, hysteresis and other piecewise-linear characteristics. This approach combines an adaptive inverse with a feedback control law. The adaptive inverse is to cancel the eect of the actuator nonlinearity, while the feedback control law can be designed as if the actuator nonlinearity were absent. For parametrizable actuator nonlinearities which have parametrizable inverses, state or output feedback adaptive inverse controllers were developed which led to linearly parametrized error models suitable for the developments of gradient projection or Lyapunov adaptive laws for updating the inverse parameters. The adaptive inverse approach can be viewed as an algorithm-based compensation approach for cancelling unknown actuator nonlinearities caused by component imperfections. As shown in this chapter, this approach can be incorporated with existing control designs such as model reference, PID, pole placement, linear quadratic, backstepping and other dynamic compensation techniques. An adaptive inverse can be added into a control system loop without the need to change a feedback control design for a known linear or nonlinear dynamics following the actuator nonlinearity. Improvements of system tracking performance by an adaptive inverse have
285
been shown by simulation results. Some signal boundedness properties of adaptive laws and closed loop systems have been established. However, despite the existence of a true inverse which completely cancels the actuator nonlinearity, an analytical proof of a tracking error convergent to zero with an adaptive inverse is still not available for a general adaptive inverse control design. Moreover, adaptive inverse control designs for systems with unknown multivariable or more general nonlinear dynamics are still open issues under investigation.
References [1] Elliott, H. and Wolovich, W. A. (1984) `Parametrization Issues In Multivariable Adaptive Control', Automatica, Vol. 20, No. 5, 533±545. [2] Goodwin, G. C. and Sin, K. S. (1984) Adaptive Filtering Prediction and Control. Prentice-Hall. [3] Hatwell, M. S., Oderkerk, B. J., Sacher, C. A. and Inbar, G. F. (1991) `The Development of a Model Reference Adaptive Controller to Control the Knee Joint of Paraplegics', IEEE Trans. Autom. Cont., Vol. 36, No. 6, 683±691. [4] Ioannou, P. A. and Sun, J. (1995) Robust Adaptive Control. Prentice-Hall. [5] Krasnoselskii, M. A. and Pokrovskii, A. V. (1983) Systems with Hysteresis. Springer-Verlag. [6] KrsticÂ, M., Kanellakopoulos, I. and KokotovicÂ, P. V. (1995) Nonlinear and Adaptive Control Design. John Wiley & Sons. [7] Merritt, H. E. (1967) Hydraulic Control Systems. John Wiley & Sons. [8] Physik Instrumente (1990) Products for Micropositioning. Catalogue 108±12, Edition E. [9] Rao, S. S. (1984) Optimization: Theory and Applications. Wiley Eastern. [10] Recker, D. (1993) `Adaptive Control of Systems Containing Piecewise Linear Nonlinearities', Ph.D. Thesis, University of Illinois, Urbana. [11] Tao, G. and Ioannou, P. A. (1992) `Stability and Robustness of Multivariable Model Reference Adaptive Control Schemes', in Advances in Robust Control Systems Techniques and Applications, Vol. 53, 99±123. Academic Press. [12] Tao, G. and KokotovicÂ, P. V. (1996) Adaptive Control of Systems with Actuator and Sensor Nonlinearities. John Wiley & Sons. [13] Tao, G. and Ling, Y. (1997) `Parameter Estimation for Coupled Multivariable Error Models', Proceedings of the 1997 American Control Conference, 1934±1938, Albuquerque, NM. [14] Tao, G. and Tian, M. (1995) `Design of Adaptive Dead-zone Inverse for Nonminimum Phase Plants', Proceedings of the 1995 American Control Conference, 2059±2063, Seattle, WA. [15] Tao, G. and Tian, M. (1995) `Discrete-time Adaptive Control of Systems with Multi-segment Piecewise-linear Nonlinearities', Proceedings of the 1995 American Control Conference, 3019±3024, Seattle, WA.
286
[16] Tian, M. and Tao, G. (1996) `Adaptive Control of a Class of Nonlinear Systems with Unknown Dead-zones', Proceedings of the 13th World Congress of IFAC, Vol. E, 209±213, San Francisco, CA. [17] Tian, M. and Tao, G. (1997) `Adaptive Dead-zone Compensation for OutputFeedback Canonical Systems', International Journal of Control, Vol. 67, No. 5, 791±812. [18] Truxal, J. G. (1958) Control Engineers' Handbook. McGraw-Hill.
11
Stable multi-input multioutput adaptive fuzzy/neural control R. OrdoÂnÄez and K. M. Passino
Abstract In this chapter, stable direct and indirect adaptive controllers are presented which use Takagi±Sugeno fuzzy systems, conventional fuzzy systems, or a class of neural networks to provide asymptotic tracking of a reference signal vector for a class of continuous time multi-input multi-output (MIMO) square nonlinear plants with poorly understood dynamics. The direct adaptive scheme allows for the inclusion of a priori knowledge about the control input in terms of exact mathematical equations or linguistics, while the indirect adaptive controller permits the explicit use of equations to represent portions of the plant dynamics. We prove that with or without such knowledge the adaptive schemes can `learn' how to control the plant, provide for bounded internal signals, and achieve asymptotically stable tracking of the reference inputs. We do not impose any initialization conditions on the controllers, and guarantee convergence of the tracking error to zero.
11.1 Introduction Fuzzy systems and neural networks-based control methodologies have emerged in recent years as a promising way to approach nonlinear control problems. Fuzzy control, in particular, has had an impact in the control community because of the simple approach it provides to use heuristic control knowledge for nonlinear control problems. However, in the more complicated situations where the plant parameters are subject to perturbations, or when the dynamics of the system are too complex to be characterized reliably by an explicit mathematical model, adaptive schemes have been introduced that
288
multi-input multi-output adaptive fuzzy/neural control
gather data from on-line operation and use adaptation heuristics to automatically determine the parameters of the controller. See, for example, the techniques in [1]±[7]; to date, no stability conditions have been provided for these approaches. Recently, several stable adaptive fuzzy control schemes have been introduced [8]±[12]. Moreover, closely related neural control approaches have been studied [13]±[18]. In the above techniques, emphasis is placed on control of single-input singleoutput (SISO) plants (except for [4], which can be readily applied to MIMO plants as it is done in [5, 6], but lacks a stability analysis). In [19], adaptive control of MIMO systems using multilayer neural networks is studied. The authors consider feedback linearizable, continuous-time systems with general relative degree, and utilize neural networks to develop an indirect adaptive scheme. These results are further studied and summarized in [20]. The scheme in [19] requires the assumptions that the tracking and neural network parameter errors are initially bounded and suciently small, and they provide convergence results for the tracking errors to ®xed neighbourhoods of the origin. In this chapter we present direct [21] and indirect [22] adaptive controllers for MIMO plants with poorly understood dynamics or plants subjected to parameter disturbances, which are based on the results in [8]. We use Takagi±Sugeno fuzzy systems or a class of neural networks with two hidden layers as the basis of our control schemes. We consider a general class of square MIMO systems decouplable via static nonlinear state feedback and obtain asymptotic convergence of the tracking errors to zero, and boundedness of the parameter errors, as well as state boundedness provided the zero dynamics of the plant are exponentially attractive. The stability results do not depend on any initialization conditions, and we allow for the inclusion in the control algorithm of a priori heuristic or mathematical knowledge about what the control input should be, in the direct case, or about the plant dynamics, in the indirect case. Note that while the indirect approach is a fairly simple extension of the corresponding single-input single-output case in [8], the direct adaptive case is not. The direct adaptive method turns out to require more restrictive assumptions than the indirect case, but is perhaps of more interest because, as far as we are aware, no other direct adaptive methodology with stability proof for the class of MIMO systems we consider here has been presented in the literature. The results in this chapter are nonlocal in the sense that they are global whenever the change of coordinates involved in the feedback linearization of the MIMO system is global. The chapter is organized as follows. In Section 11.2 we introduce the MIMO direct adaptive controller and give a proof of the stability results. In Section 11.3 we outline the MIMO indirect adaptive controller, giving just a short sketch of the proof, since it is a relatively simple extension of the results in [8]. In Section 11.4 we present simulation results of the direct adaptive method
Adaptive Control Systems
289
applied, ®rst, to a nonlinear dierential equation that satis®es all controller assumptions, as an illustration of the method, and then to a two-link robot. The robot is an interesting practical application, and it is of special interest here because it does not satisfy all assumptions of the controller; however, we show how the method can be made to work in spite of this fact. In Section 11.5 we provide the concluding remarks.
11.2 Direct adaptive control Consider the MIMO square nonlinear plant (i.e. a plant with as many inputs as outputs [23, 24]) given by X_ f X g1 Xu1 gp Xup y1 h1 X
111
yp hp X
where X x1 xn T R n is the state vector, U : u1 up T R p is the control input vector, Y : y1 yp T R p is the output vector, and f gi hi i 1 p are smooth functions. If the system is feedback linearizable [24] by static state feedback and has a well-de®ned vector relative degree r : r1 rp T , where the ri 's are the smallest integers such that at least one of r the inputs appears in y i i , input±output dierential equations of the system are given by ri
yi
L rj i hi
p j1
Lgj L fri 1 hi uj
112
with at least one of the Lgj L fri 1 hi 0 (note that Lf hX : R n R is the Lie @h derivative of h with respect to f , given by Lf hX @X f X. De®ne, for ri ri 1 convenience, i X : L f hi and i j X : Lgj Lf hi . In this way, we may rewrite the plant's input±output equation as r1 11 1p u1 1 y1 11:3 rp up p p1 pp yp Y r t
AX;t
BX;t
Ut
Consider the ideal feedback linearizing control law, U u 1 ; ; u p T ,
290
Stable multi-input multi-output adaptive fuzzy/neural control
U B 1 A m
114
(note that, for convenience, we are dropping the references to the independent variables except where clari®cation is required), where the term m 1 ; ; p T is an input to the linearized plant dynamics. In order for U to be well de®ned, we need the following assumption: (P1) Plant Assumption The matrix B as de®ned above is nonsingular, i.e. B 1 exists and has bounded norm for all X Sx ; t 0, where Sx R n is some compact set of allowable state trajectories. This is equivalent to assuming p B min > 0 B2 1 B max <
11:5 11:6
where p B and 1 B are, respectively, the smallest and largest singular values of B. In addition, in order to be able to guarantee state boundedness under state feedback linearization, we require: (P2) Plant Assumption The plant is feedback linearizable by static state feedback; it has a general vector relative degree r r1 ; ; rp T , and its zero dynamics are exponentially attractive (please refer to [24] for a review on the concept of zero dynamics and static state feedback of square MIMO systems). We also assume the state vector X to be available for measurement. Our goal is to identify the unknown control function (11.4) using fuzzy systems. Here we will use generalized Takagi±Sugeno (T±S) fuzzy systems with centre average defuzzi®cation. To brie¯y present the notation, take a fuzzy system denoted by f~X; W (in our context, X could be thought of as the state vector, and W as a vector of possibly exogenous signals). Then,
R ci i ~ f X; W i1 . Here, singleton fuzzi®cation of the input vectors R i1 i X x1 ; . . . ; xn T , W w1 ; . . . ; wq T is assumed; the fuzzy system has R rules, and i is the value of the membership function for the premise of the ith rule given the inputs X, W. It is assumed that the fuzzy system is R n constructed in such a way that 0 i 1 and i1 i 0 for all X R , q W R . The parameter ci is the consequent of the ith rule, which in this chapter will be taken as a linear combination of Lipschitz continuous functions, zk X R; k 1; . . . ; m 1, so that ci ai;0 ai;1 z1 X ai;m1 zm1 X; i 1; . . . ; R. De®ne
Adaptive Control Systems
z
1 z1 X .. . zm1 X
Rm;
T
a1;0
a1;1
a2;0 A .. .
a2;1 .. .
.. .
aR;1
aR;m1
T
aR;0
a1;m1
291
1 X; W; . . . ; R X; W
R ; i1 i X; W
a2;m1 .. .
Then, the nonlinear equation that describes the fuzzy system can be written as f~X; W z T XAX; W (notice that standard fuzzy systems may be treated as special cases of this more general representation). It was shown in [8] that the T±S model can represent a class of two-layer neural networks and many standard fuzzy systems. Note that while may depend on both X and W and is bounded for any value they may take, z depends on X only. This allows us to impose no restrictions on W to guarantee boundedness of the fuzzy system. We will represent the ith component of the ideal control (11.4), i 1; . . . ; p, as
where
A i
u i X; W; t z Ti XA i i X; W di X; W uki t R
mi Ri
is assumed to exist, and is de®ned by T z i Ai i uki u i A i : arg min sup Ai Oi XSx ;WSw ;t 0
11:7 11:8
and the representation error using optimal parameters (which arises because of the ®nite number of basis functions used) is bounded, i.e. di X; W Di X; W, where Di is a known bounding function. We de®ne Oi as a compact set within which the matrix of coecients estimates, Ai t, is allowed to lie, Sx R n as the subspace through which the state trajectory may travel under closed loop control, and Sw R q is the subset where the vector W may lie (notice that we do not restrict the sizes of Sx and Sw ; however, the sup in (11.8) is assumed to exist). As a result of the proof we will be able to determine that X actually remains within a compact subset of Sx . Note that the ideal control law (11.4) is a function not only of the states, but also of m, which may depend on variables other than the states (as will be described below). The vector W is provided to account for this dependence. The term uki represents a known part of the ideal control input, which may be available to the designer through knowledge of the plant or expertise. If it is not available, this term may be set equal to zero with all the properties of the adaptive controller still holding. The only restriction on uki is that it must be bounded. Note that an appropriate use of uki may help to
292
Stable multi-input multi-output adaptive fuzzy/neural control
signi®cantly improve the performance of the controller, even though in principle it has no eect on stability. Thus, the fuzzy system approximation of u i is given by u^i X W t : z Ti XAi ti X W uki t
119
The matrix Ai t is to be adjusted adaptively on-line in order to try to improve the approximation. We de®ne the parameter error matrix, Fi t : Ai t A i . ^ : ^ Let U u1 . . . u^p T . Our objective is to have the plant outputs track a vector of reference trajectories, Ym ym1 . . . ymp T , on which we make the following assumption: (R1) Reference Input Assumption The desired reference trajectories ymi are ri times continuously dierentiable, r with ymi . . . y mii measurable and bounded, for i 1 . . . p. We de®ne the output errors eoi : ymi yi . De®ne also the error signals esi : k0 . . . k iri 2 1eoi . . . e ori i 2 e ori i 1 T and
i i ri 1 T i esi : e_si e r oi k 0 . . . k ri 2 e_ oi . . . e oi
The coecients of 1 1 : r 1 s i k iri 2 s ri 2 . . . k i1 s k i0 L^i s i 1 . . . p, are picked so that the transfer functions are stable. Let the ith r component of the parameter m in (11.4) be given by i : y mii i esi esi , where i > 0 is a constant. Consider the control law ^ Ud UU
11:10
where Ud : ud1 ; . . . ; udp T is a control term required to ensure stability that will be de®ned later. From (11.3) and (11.4) we can derive an expression for the plant output Y r A BU A BU U BU m BU U
11:11
Then, using the previous de®nitions, the ith component of the output error dynamics is given by ri
ri e ori i y m yi i
so that
i y r m i i
p j1
i j uj u j i esi esi
p j1
i j uj u j 11:12
Adaptive Control Systems
e_si i esi
p j1
i j uj u j
293
11:13
Before proceeding we need to introduce another set of assumptions on the ^ plant and formalize our assumptions about the control term U. (P3) Plant Assumption Each entry of B, besides those on the main diagonal, is bounded by known constants, i j X i j ; i; j 1; . . . ; p; i j. We require that the entries in the main diagonal satisfy 0 < ii ii X ii < ; i 1; . . . ; p, and their derivatives be de®ned and satisfy _ii X Mii X; i 1; . . . ; p, where ii , ii and Mii X are known bounds. Furthermore, the bounds have to satisfy p 1 i j < 1; ii j1;j i
i 1; . . . ; p:
11:14
(C1) Direct Adaptive Control Assumption Bounding functions Ui X; W such that z Ti Xti X; W Ui X; W; i 1; . . . ; p; X Sx ; W Sw , are known and they are continuous functions. Furthermore, the fuzzy systems or neural networks that de®ne the control ^ are de®ned so that the bounding functions of the representation term U, errors, Di X; W; i 1; . . . ; p, are continuous. Note that in (P3) the entries of the main diagonal of B are all assumed positive. This is only to simplify the analysis; the diagonal entries may have any sign, as long as they are bounded away from zero, and the stability analysis requires only slight modi®cations to accommodate such a case. In (C1), knowledge of the bounding function Ui X; W is reasonable, since both zi X and i X; W are known: a projection algorithm may be employed to guarantee that Ai t stays within the compact set Oi of allowable parameters. Then, an upper estimate of i t can be computed, and Ui can be de®ned. Consider the function Vi
1 2 1 e tr Ti Qui i 2 ii si 2
11:15
with Qui R mi mi positive de®nite and diagonal. This function quanti®es both the tracking error for the ith plant output and the approximation error for the parameter estimates of the ith term of (11.4). Taking the derivative of (11.15) yields
294
Stable multi-input multi-output adaptive fuzzy/neural control
1 _ii V_ i esi e_si tr Ti Qui _ i 2 e 2si ii 2 ii
11:16
p 1 _ii i j uj u j tr Ti Qui _ i 2 e 2si esi i esi ii 2 ii j1
11:17
De®ne the adaptation law for the T±S fuzzy system or neural network as T A_ i : Q 1 ui zi i esi
11:18
so that, applying the properties of the trace operator and the fact that _ i A_ i , we obtain tr Ti Qui _ i z Ti i i esi . Noting that ui u i udi z Ti i i di , we get p p p e _ii i s i T 2 i j z j j j 2 e 2si 11:19 i j dj i j udj V_ i e si ii ii 2 ii j1 j1 j1;j i p p i j _ii i j i V_ e 2si esi udi esi Dj udj z Tj j j 2 e 2si ii 2 ii j1 ii j1;j i ii 11:20 De®ne
p i j Uj j1 ii j1;j i ii Mii X i : esi : 2 2ii
i :
p i j
Dj
Given these de®nitions, let
p i j Umax udi : sgnesi i j1;j 1 ii
i
11:21
where we need to derive an expression for Umax such that udi Umax ; i 1; . . . ; p. From (11.21) we have udi i i Umax
p i j Umax j1;j i ii
11:22
It follows that if we choose
i i Umax t max
p i j i1;...;p 1 ji;j 1 ii
11:23
Adaptive Control Systems
295
and if (P3) holds, then in fact udi Umax i 1 . . . p. Using (11.21) we can now establish i 11:24 V_ i e 2si ii We are now ready to present our main result and give its proof. Theorem 2.1 Stability and tracking results using MIMO direct adaptive control If the reference input assumption (R1) holds, the plant assumptions (P1), (P2) and (P3) hold, and the control law is de®ned by (11.10) with the control assumption (C1) and the adaptive laws (11.18) are used, Then the following holds: r 1
(1) The plant states, as well as its outputs and their derivatives, yi ; . . . ; y i i , i 1; . . . ; p, are bounded. (2) The control signals are bounded, i.e. U t : supt t < . (3) The magnitudes of the output errors, eoi , decrease at least asymptotically to zero, i.e. limt eoi 0; i 1; . . . ; p. Proof
To show part 1, consider the Lyapunov candidate V :
p
Vi
i1
The above analysis guarantees V_
p i i1
ii
e 2si
11:25
11:26
so V is a positive de®nite function with negative semide®nite derivative. This implies that Vi ; therefore, esi ; esi and i for i 1; . . . ; p (notice that this analysis alone does not guarantee Ai i for all time; rather, a projection algorithm should be used to achieve this). From the de®nition of esi j we have e oi G^ ji sesi , where G^ ji s : s j =L^i s; j 0; . . . ; ri 1. Since, by j de®nition, G^ ji is stable, e oi ; j 0; . . . ; ri 1, and since by assumption ri r 1 (R1) the signals y mi are bounded, we conclude that yi ; . . . ; y i i ; i 1; . . . ; p. With the outputs bounded, and together with assumption (P2) we have that the states x1 ; . . . ; xn are bounded [23], which implies that the state trajectories are limited to a bounded subset of Sx . Let Sx be the compact ball of minimum radius that contains the bounded subset of state trajectories. Since i is continuous and zi is Lipschitz continuous by de®nition in Sx , then they are uniformly continuous, and therefore bounded, on Sx , and given that uki is bounded, we have u^i ; i 1; . . . ; p. Since Ui is de®ned as a continuous
296
Stable multi-input multi-output adaptive fuzzy/neural control
function, and Di is assumed continuous for all X Sx W Sw , both are bounded on Sx , so i , and i because esi from part 1. This implies Umax , so udi ; i 1; . . . ; p by construction. Hence, U . To prove part 3 we notice that, from (11.26) p i 2 _ Vdt 11:27 e si dt 0 0 i1 ii V0 V <
11:28
This establishes that esi 2 ; i 1; . . . ; p 2 t : 0 2 t dt < . Having determined that e joi ; j 1; . . . ; ri 1, it follows that e_si , so esi is uniformly continuous. Since esi 2 and e_si , by Barbalat's lemma we have asymptotic stability of esi (i.e. limt esi 0), which implies asymptotic stability of eoi (i.e. limt eoi 0), for i 1; . . . ; p. Notice that although assumption (P1) is not used explicitly in the proof, it is still necessary in order to guarantee the existence of U , without which the argument is not sound. Remark 2.1 Note that, although in principle the choice of the vector W is arbitrary, a typical choice may be an error vector, i.e. W Y Ym , or some other function of the plant outputs and the reference model outputs. In this way, as a result of the proof, we also obtain that W remains within a bounded subset of Sw .
11.3
Indirect adaptive control
Here we consider, again, the class of plants de®ned in (11.1). If assumptions (P1) and (P2) of Section 11.2 are satis®ed, then we may rewrite the input± output form of the plant as r1 11 11k 1p 1pk 1 1k u1 y1 . .. .. .. .. .. .. 11:29 . . . . . r up p pk p1 p1k pp ppk yp p AX;t
BX;t
Ut
where ik t and i jk t are known components of the plant's dynamics (that may depend on the state) or exogenous time dependent signals, with the only constraint that they have to be bounded for all t 0. Throughout the following analysis they may be set equal to zero for all t; however, as in the direct adaptive case, a good choice of these known functions may help improve the performance of the controller. The functions i X and i j X represent unknown nonlinear dynamics of the plant.
Adaptive Control Systems
297
Again consider the ideal feedback linearizing control law (11.4), where the term m will be rede®ned below. Our goal is to identify the unknown functions i and i j using fuzzy systems (or neural networks) in order to indirectly approximate the ideal control law U . Let the fuzzy system be a Takagi± Sugeno (T±S) form with centre average defuzzi®cation as in Section 11.2. We rewrite i X z Ti A i i di X
11:30
i j X z T i j A i j i j d i j X
11:31
where A i R mi pi and A i j R m i j p i j are de®ned by A i : arg min sup z Ti Ai i i Ai i XSx
A i j : arg min
sup z T i j A i j i j i j
A i j i j XSx
11:32
11:33
Note that we are assuming the ability to specify fuzzy systems in such a way that the representation errors using optimal parameters (which arise because of the ®nite number of basis functions used) are bounded, i.e. di X Di X; d i j X D i j X, where Di X and D i j X are known bounding functions. We require the representation errors D i j X to be small (later we will provide an explicit condition as to how small they have to be), which means that the matrix B can, ideally, be well approximated by our fuzzy systems (or neural networks) using optimal parameters. Our adaptive controller's stability will not, however, depend on its ability to identify these optimal parameters. The compact set Sx R n is de®ned as before, and i ; i j are compact sets within which the parameter matrices estimates, Ai t and A i j t, are allowed to lie. Thus, the fuzzy system approximations of i X and i j X are given by ^i X : z Ti Ai i
11:34
^i j X : z T i j A i j i j
11:35
The matrices Ai t and A i j t are to be adjusted adaptively on line in order to try to improve the approximation. We de®ne the output errors eoi and error signals esi and esi as in Section 11.2. Consider the control law U : Uce , where Uce uce1 ; . . . ; ucep T is a certainty ^ : ^i j X ij t, equivalence control term. De®ne the matrix B k ^ i; j 1; . . . ; p. B is an approximation of the ideal and unknown matrix B. ^ 1 , i; j 1; . . . ; p be a matrix of the elements of the Furthermore, let bi j : B ^ 1 exists for all X Sx and t 0. If the sets inverse. We need to ensure that B
i j are constructed such that
298
Stable multi-input multi-output adaptive fuzzy/neural control
^ min p B
11:36
^ max 1 B
11:37
for all X Sx , then, as long as the matrices A i j remain within i j , ^ 1 exists (this can be achieved using a respectively, we can guarantee that B projection algorithm). Note that if we knew the matrix B to be, for instance, strictly diagonally dominant (as required by the Levy±Desplanques theorem (see [25] for an explanation of this and other invertibility results)) with known lower bounds for the main diagonal entries, we could relax the conditions on ^ in a strictly the sets i j by applying instead a projection algorithm that kept B diagonally dominant form similar to B to ensure it is invertible. In order to cancel the unknown parameter errors we use the following adaptive laws: A_ i Q 1 i zi i esi
11:38
A_ i j Q 1 i j z i j i j esi ucej with Qi R mi mi ; Q i j R m i j m i j positive de®nite and diagonal. Now we can write an expression for Uce ^ m ^ 1 A Uce : B
11:39
^ : ^1 1k ; . . . ; ^p pk T . Here mt 1 t; . . . ; p t T is chosen where A to provide stable tracking and to allow for robustness to parameter uncertainty. Namely, let i si Di X sgnesi Umax X sgnesi i t : y r mi i esi e
p j1
D i j X 11:40
where i > 0 is a design parameter, and Umax X is a function chosen so that ucei Umax X; i 1; . . . ; p. It can be shown that the choice ai X Umax X max 11:41 i1;...;p 1 ci X
r ^j jk y mjj j esj satis®es the requirement, where ai X : pj1 bi j
esj Dj and ci X : pj1 bi j pl1 D j l . It should be noted that for many classes of plants, each i j is a smooth function easily represented by a fuzzy system. For example, if each i j may be expressed as a constant, then D i j 0 for all i; j since a fuzzy system may exactly represent a constant on a compact set. This would remove the need for the Umax X term to be included in (11.40). At this point we need to formalize our general assumption about the controller:
Adaptive Control Systems
299
(C2) Indirect Adaptive Control Assumption The fuzzy systems, or neural networks, that de®ne the approximations (11.34) and (11.35), are de®ned so that Di X ; D i j X , for all X Sx R n ; i; j 1; . . . ; p. Furthermore, the ideal representation errors D i j X are small enough, so that we have 0 ci X < 1, for all X Sx ; i 1; . . . ; p. Notice that the maximum sizes of D i j that satisfy (C2) can be found, since ^ 1 1 , and as long as assumption (C2) from (11.36) we have bi j X B 2 min is satis®ed, we ensure that ucei Umax X; i 1; . . . ; p as desired. We have completely speci®ed the signals that compose the control vector U, and now we state our main result; its proof is omitted, but follows ideas used in Section 11.2 and [8]. Theorem 3.1 Stability and tracking results using MIMO indirect adaptive control If the reference input assumption (R1) holds, the plant assumptions (P1) and (P2) hold, and the control law is de®ned by (11.39) with the control assumption (C2), Then the following holds: (1) The plant states, as well as its outputs and their derivatives, yi ; . . . ; yi ri 1 i; i 1; . . . ; p, are bounded. (2) The control signals are bounded, i.e. ucei ; i 1; . . . ; p. (3) The magnitudes of the output errors, eoi , decrease at least asymptotically to zero, i.e. limt eoi 0; i 1; . . . ; p.
11.4 Applications 11.4.1 Illustrative example Consider the nonlinear dierential equation given by x_ 1 x2 0 3u1 u2 x_ 2 x1 x 22 x3 x_ 3 x1 2x2 3x3 x1 u1 22 0:5 sinx1 u2
11:42
Notice that these are coupled nonlinear dynamics. The B matrix is not constant, but contains a bounded function of the states. We are interested in the outputs y1 x1 and y2 x3 . It is easily veri®ed that this system has a vector relative degree of 2; 1 T . We de®ne the error equations as es1 eo1 e_o1 and es2 eo2 , with the output errors de®ned appropriately. We want the outputs of the system to track the reference vector Ym1 s; Ym2 s T T R1 s R2 s ; , where R1 s r1 t and R2 s r2 t E is the s1 2 s1 Laplace transform operator). Thus, e_o1 is computed from y_ m1 and x2 .
300
Stable multi-input multi-output adaptive fuzzy/neural control
Table 11.1 Rule base
F 21 F 22 F 23 F 24 F 25 F 26 F 27 F 28 F 29
F 11
F 12
F 13
F 14
F 15
F 16
F 17
F 18
F 19
c k1 c k1 c k1 c k1 c k1 c k2 c k3 c k4 c k5
c k1 c k1 c k1 c k1 c k2 c k3 c k4 c k5 c k6
c k1 c k1 c k1 c k2 c k3 c k4 c k5 c k6 c k7
c k1 c k1 c k2 c k3 c k4 c k5 c k6 c k7 c k8
c k1 c k2 c k3 c k4 c k5 c k6 c k7 c k8 c k9
c k2 c k3 c k4 c k5 c k6 c k7 c k8 c k9 c k9
c k3 c k4 c k5 c k6 c k7 c k8 c k9 c k9 c k9
c k4 c k5 c k6 c k7 c k8 c k9 c k9 c k9 c k9
c k5 c k6 c k7 c k8 c k9 c k9 c k9 c k9 c k9
We use two T±S fuzzy systems to produce u^1 and u^2 , and we set the `known' controller terms uki 0 i 1 2. Both fuzzy systems have eo1 and eo2 as their inputs (so here W eo1 eo2 T , and we let z Tk 1 x1 x2 x3 k 1 2. Both fuzzy systems have nine triangular membership functions for each of the two input universes of discourse, uniformly distributed over the interval 1 1 with 50% overlap (we use scaling gains to normalize the inputs to this interval). We saturate the outermost membership functions, and the output is computed using centre average defuzzi®cation. Both systems' coecient matrices, A1 and A2 , are initialized with zeroes, and they utilize the rule base shown in Table 11.1. The labels F ji denote the ith fuzzy set for the jth input, where i 1 corresponds to the leftmost, and i 9 to the rightmost fuzzy set. Each entry of the table corresponds to one output function c ki i 1 . . . 9, where c ki z Tk a ki , and a ki is the ith column of Ak k 1 2. As an example, consider the rule for c k3 that is inside a box in the table If eo1 is F 13 and eo2 is F 25 then c k3 z Tk a k3 where we evaluate the and operation using minimum. Note that i i 1 . . . 9, is the result of evaluating the premise of the ith rule. From the plant's equation we choose the bounds 11 3:3; 11 2:7; 22 5:3; 22 2:7; 21 1:3; 12 1:3; M11 0:0, and M22 x2 (these two bounds are obtained by dierentiating the diagonal entries of the matrix B). Also, the fuzzy system approximation error bounds are chosen as D1 0:1; D2 0:1 (note that this choice is not readily apparent from the de®nitions of the fuzzy systems; rather, the bounds are found through a trial and error procedure). Since we know the vectors z Tk and Tk ; k 1; 2, an easy way to compute the bounding functions U1 and U2 is to use fuzzy systems that have the same structure as the ones used for u^1 and u^2 . Initially we chose the
Adaptive Control Systems
301
entries in their coecient matrices to be large, so that they would bound the values taken by the u^1 and u^2 fuzzy systems. We found, however, that this created high amplitude and high frequency oscillations of the control signals (due to the term Ud ), which are undesirable. After some tuning we determined that setting the U1 and U2 fuzzy system's coecients to 0.1 gave us stable behaviour, good tracking and much smoother control signals. Finally, the adaptation gains were chosen as Qu1 Qu2 24I, where I is a 2 2 identity matrix. In Figure 11.1 we observe the results for direct adaptive control on this system. We used a fourth order Runge±Kutta numerical approximation to the dierential equation solution, with a step size of 0.001. The reference inputs r1 t and r2 t are chosen as square waves, and the corresponding reference model outputs are plotted in Figure 11.1(a) with dashed lines, but are hard to see since x1 and x3 track them closely. In Figure 11.1(b) we observe the applied control inputs.
11.4.2 Application to a two-link robot arm Next, we consider direct adaptive control of a two-degree of freedom robot arm. This system does not satisfy assumption (P3) because, as we will see, the matrix multiplying the input vector U contains functions of the states (similar to the example in the previous section). However, in some regions the bounds for the entries do not satisfy the diagonal dominance condition. Nevertheless, our simulation results show that the method seems to be able to provide stable tracking with adequate performance; furthermore, the controller is able to compensate for an `unknown' change in system parameters, which represents the situation where the robot picks up an object after some time of nominal operation (i.e. when the robot is not holding any object). The robot arm consists of two links, the ®rst one mounted on a rigid base by means of a frictionless hinge, and the second mounted at the end of link one by means of a frictionless ball bearing. The inputs to the system are the torques 1 and 2 applied at the joints. The outputs are the joint angles 1 and 2 . A mathematical model of this system can be derived using Lagrangian equations, and is given by
h_2 1 H21 H22 2 h_1 H11
H12
H
where
h_1 h_2 0
_ 1 1 g 1 _2 g2 2
11:43
302
Stable multi-input multi-output adaptive fuzzy/neural control X1 0
−5
−10 0
20
40
60
80
100 X2
120
140
160
180
200
20
40
60
80
100 X3
120
140
160
180
200
20
40
60
80
100 Time (s)
120
140
160
180
200
120
140
160
180
200
120
140
160
180
200
2
0
−2 0 10
5
0 0
(a) U1 0 −5 −10 −15 −20 0
20
40
60
80
100 U2
50 40 30 20 10 0 0
20
40
60
80
100 Time (s)
(b) Figure 11.1 (a) System states (solid lines) and reference model outputs (dashed lines). (b) Control inputs
Adaptive Control Systems
303
H11 I1 I2 m1 l 2c1 m2 l 21 l 2c2 2l1 lc2 cos2 m3 l 21 l 22 2l1 l2 cos2 H22 I2 m2 l 2c2 m3 l 22 H12 H21 I2 m2 l 2c2 l1 lc2 cos2 m3 l 22 l1 l2 cos2 h m2 l1 lc2 sin2 g1 m1 lc1 g cos1 m2 glc2 cos1 2 l1 cos1 g2 m2 lc2 g cos1 2 The matrix H can be shown to be positive de®nite, and therefore always invertible. In our simulation we use the following parameter values: m1 1:0 kg, mass of link one; m2 1:0 kg, mass of link two; l1 1:0 m, length of link one; l2 1:0 m, length of link two; lc1 0:5 m, distance from the joint of link one to its centre of gravity; lc2 0:5 m, distance from the joint of link two to its centre of gravity; I1 0:2 kg m 2 , lengthwise centroidal inertia of link one; and I2 0:2 kg m 2 , lengthwise centroidal inertia of link two. The mass m3 , initially set equal to zero, represents the mass of an object at the end of the link. After 100 seconds of operation, the robot `picks up' an object of mass m3 3:0 kg. We can rewrite the system dynamics in input output form H22 h_2 2_1 _2 H12 h_ 21 H22 g1 H12 g2 1 1 H11 H22 H12 H21 H21 h_2 2_1 _2 H11 h_ 21 H21 g1 H11 g2 2 H22 H12 1 2 H21 H11
We observe that the input vector is multiplied by H 1 , which contains the function cosg2 . For some values of 2 , H 1 is not diagonally dominant, and thus does not satisfy assumption (P3) (note, e.g., that for some 2 values, H22 < H12 . However, we found that not only was it possible to make the direct adaptive method work, but also that is oered relatively good performance and the ability to handle system parameter changes. We would like the outputs 1 and 2 to track desired reference angles, which are obtained from the reference model vector Ym1 s; Ym2 s T T 0:75 2 R1 s 1:5 2 R2 s , where R1 s r1 t and R2 s r2 t. ; s 0:75 2 s 1:5 2 Clearly, the system has a vector relative degree 2; 2 T , so, letting eo1 ym1 1 and eo2 ym2 2 , we de®ne the error equations es1 eo1 e_o1 and es2 eo2 e_o2 . The error derivatives are available, since the reference inputs are twice dierentiable, and the angle derivatives _1 and _2 are plant states. When designing our fuzzy systems we assumed that there is no strong cross-
304
Stable multi-input multi-output adaptive fuzzy/neural control
coupling between the inputs and outputs, which greatly simpli®es the design: we chose eo1 and e_o1 as inputs to the fuzzy system for 1 (so for this fuzzy system W eo1 ; e_o1 T , and eo2 and e_o2 as inputs for 2 . Both fuzzy systems have z Ti 1; _1 ; 1 ; _2 ; 2 , and are otherwise structurally identical to the systems de®ned in Section 11.4.1. Here, also, we initialized the coecient matrices of both systems with zeroes. Since, as mentioned before, the system does not satisfy assumption (P3) for all 2 , the way to choose the bounds for the H 1 matrix entries is not clear. In view of this we took a pragmatic approach, where we ®rst picked bounds that were as close as possible to the real bounds, and yet satis®ed assumption (P3): letting B H 1 , substitution of the numerical values of the parameters shows that, taking into account both values of m3 ; 1:1 < 11 < 1:2; 2:3 < 12 ; 21 < 2:5, and 0:7 < 22 < 7:3. Thus, we picked 11 1:3; 11 1:1; 22 2:3; 22 2:5; 21 2:3, and 12 2:3. This choice resulted in somewhat acceptable behaviour, but with highly oscillatory control signals. Therefore, we decided to tune these bounds, even though the theoretical assumptions were violated. We found that reducing the size of the bounds while meeting the diagonal dominance condition yielded satisfactory results: the magnitude of the control signals' oscillations was drastically reduced, and at the same time we obtained adequate tracking and apparent robustness to the plant parameter change we investigated. We ®nally chose the and 21 12 0:3. bounds as 11 11 1:2; 22 22 1:2, Dierentation of the diagonal entries of B yields M11 0:0, and M22 3:1_2 . We picked the fuzzy system approximation error bounds as D1 0:1; D2 0:1 (again this is the result of a tuning process). The bounding functions U1 and U2 were picked in a way similar to Section 11.4.1, where the matrix coecients are ®rst chosen large, and then decreased until adequate performance is achieved; setting the coecients to 0.1 gave us the best results. Finally, the adaptation gains were set to Qu1 Qu2 4:71I, where I is a 2 2 identity matrix. We observe the control results on Figure 11.2. We let r1 t and r2 t be square waves. Initially the controller has some diculties, and tracking is not perfect: at this point, the T±S coecient matrices are moving away from zero and possibly adapting towards values that allow for better tracking. After the ®rst period of the square wave reference inputs we note that tracking improves signi®cantly. At time t 100 seconds, when the system dynamics change as the robot `picks up' an object, we ®nd the outputs exhibit virtually no transient overshoot, and tracking continues to be adequate. However, at this point we observe a high peak in 2 , as the controller tries to compensate for the increase in the load of link two. The peaks recur at the transition points, where the references step up or down, but they tend to decrease in magnitude (we let the simulation run for 12 000 seconds, and found this pattern to hold).
Adaptive Control Systems Theta_1
Theta_2
4
1
3
0.5
2
0
1
−0.5
0
−1
−1 0
100
305
200
300
−1.5 0
100
Theta_3
200
300
Theta_4
1
3
0
2 1
−1
0 −2
−1
−3 −4 0
−2 100
200
−3 0
300
100
200
300
(a) Tau_1 30 20 10 0 −10 −20 0
50
100
150
200
250
300
200
250
300
Tau_2 20 15 10 5 0 −5 −10 0
50
100
150
(b) Figure 11.2 (a) System states (solid lines) and reference model outputs (dashed lines). (b) Control inputs
306
11.5
Stable multi-input multi-output adaptive fuzzy/neural control
Conclusions
In this chapter we have developed direct and indirect adaptive MIMO control schemes which use Takagi±Sugeno fuzzy systems or a class of neural networks. We have proven stability of the methods and shown that they guarantee asymptotic convergence of the tracking errors to zero, as well as boundedness of all the signals and parameter errors, regardless of any initialization constraints. Both methods allow for the inclusion of previous knowledge or expertise in form of linguistics regarding what the control input should be, in the direct case, or what the plant dynamics are, in the indirect case. We show that with or without such knowledge the stability and tracking properties of the controllers hold, and present two simulations for direct adaptive control that illustrate the method.
References [1] Procyk, T. and Mamdani, E. (1979). `A Linguistic Self-organizing Process Controller' Automatica, 15(1), 15±30. [2] Driankov, D., Hellendoorn, H. and Reinfrank, M. M. (1993). An Introduction to Fuzzy Control. Springer-Verlag, Berlin Heidelberg. [3] Layne, J. R., Passino K. M. and Yurkovich, S. (1993). `Fuzzy Learning Control for Anti-skid Braking Systems', IEEE Trans. Control Systems Tech., 1(2) 122±129, June. [4] Layne, J. R., and Passino, K. M. (1993). `Fuzzy Model Reference Learning Control for Cargo Ship Steering', IEEE Control Systems Magazine, 13(6), 23±34, Dec. [5] Kwong, W. A., Passino, K. M., Lauknonen, E. G. and Yurkovich, S. (1995). `Expert Supervision of Fuzzy Learning Systems for Fault Tolerant Aircraft Control', Proc. of the IEEE, Special Issue on Fuzzy Logic in Engineering Applications, 83(3) 466±483, March. [6] Moudgal, V. G., Kwong, W. A., Passino K. M. and Yurkovich, S. (1995). `Fuzzy Learning Control for a Flexible-link Robot', IEEE Transactions on Fuzzy Systems, 3(2), 199±210, May. [7] Kwong, W. A. and Passino, K. M. (1996). `Dynamically Focused Fuzzy Learning Control', IEEE Trans. on Systems, Man, and Cybernetics, 26(1) 53±74, Feb. [8] Spooner, J. T. and Passino, K. M. (1996). `Stable Adaptive Control using Fuzzy Systems and Neural Networks', IEEE Transactions in Fuzzy Systems, 4(3), 339±359, August. [9] Wang, Li-Xin. (1994). Adaptive Fuzzy Systems and Control: Design and Stability Analysis. Prentice-Hall, Englewood Clis, NJ. [10] Wang, Li-Xin. (1992). `Stable Adaptive Fuzzy Control of Nonlinear Systems', in Proc. of 31st Conf. Decision Contr., 2511±2516, Tucson, Arizona. [11] Su, Chun-Yi. and Stepanenko, Y. (1994). `Adaptive Control of a Class of Nonlinear Systems with Fuzzy Logic', IEEE Trans. Fuzzy Systems, 2(4), 285± 294, November.
Adaptive Control Systems
307
[12] Johansen, T. A. (1994). `Fuzzy Model Based Control: Stability, Robustness, and Performance Issues', IEEE Trans. Fuzzy Systems, 2(3), 221±234, August. [13] Nerendra, K. S. and Parthasarathy, K. (1990). `Identi®cation and Control of Dynamical Systems using Neural Networks', IEEE Trans. Neural Networks, 1(1), 4±27. [14] Polycarpou, M. M. and Ioannou, P. A. (1991). `Identi®cation and Control of Nonlinear Systems Using Neural Network Models: Design and Stability Analysis. Electrical Engineering ± Systems Report 91-09-01, University of Southern California, September. [15] Sanner, R. M. and Jean-Jacques E. Slotine (1992). `Gaussian Networks for Direct Adaptive Control', IEEE Trans. Neural Networks, 3(6), 837±863. [16] Yes° ildirek, A. and Lewis, F. L. (1994). `A Neural Network Controller for Feedback Linearization', in Proc. of 33rd Conf. Decision Contr., 2494±2499, Lake Buena Vista, FL, December. [17] Chen, F.-C. and Khalil, H. K. (1992). Adaptive Control of Nonlinear Systems Using Neural Networks' Int. J. Control, 55(6), 1299±1317. [18] Rovithakis, G. A. and Christodoulou, M. A. (1994). `Adaptive Control of Unknown Plants Using Dynamical Neural Networks', IEEE Trans. Syst. Man, Cybern., 24(3), 400±412, March. [19] Liu, Chen-Chung and Chen, Fu-Chuang (1993). `Adaptive Control of Non-linear Continuous-time Systems using Neural Networks ± General Relative Degree and MIMO Cases', International Journal of Control, 58(2), 317±355. [20] Chen, Fu-Chuang and Khalil, H. K. (1995). `Adaptive Control of a Class of Nonlinear Systems using Neural Networks', in 34th IEEE Conference on Decision and Control Proceedings, New Orleans, LA, 2427±2432. [21] Ordonez, R. E. and Passino, K. M. (1997). `Stable Multi-input Multi-output Direct Adoptive Fuzzy Control', in Proceedings of the American Control Conference, 1271±1272, Albuquerque, NM, June. [22] Ordonez, R. E. Spooner, J. T. and Passino, K. M. (1996). `Stable Multi-input Multi-output Adaptive Fuzzy Control', In IEEE Conference on Decision and Control, 610±615, Kobe, Japan, September. [23] Shankar Sastry, S. and Bodson, M. (1989). Adaptive Control: Stability, Convergence, and Robustness, Prentice-Hall, Englewood Clis, New Jersey. [24] Shankar Sastry, S. and Isidori, A. (1989). `Adaptive Control of Linearizable Systems', IEEE Transac. Autom. Contr., 34(11), 1123±1131, November. [25] Horn, R. A. and Johnson, C. R. (1985). Matrix Analysis. Cambridge University Press, Cambridge (Cambridgeshire); New York.
12
Adaptive
J.-X. Xu, Q.-W. Jia and T.-H. Lee
Abstract This chapter presents a new adaptive robust control scheme for a class of nonlinear uncertain dynamical systems. To reduce the robust control gain and widen the application scope of adaptive techniques, the system uncertainties are classi®ed into two dierent categories: the structured and non-structured uncertainties. The structured uncertainty can be separated and expressed as the product of known functions of states and a set of unknown constants. The upper bounding functions of the non structured uncertainties to be addressed in this chapter is only partially known with unknown parameters. Moreover, the bounding function is convex to the set of unknown parameters, i.e. the bounding function is no longer linear in parameters. The structured uncertainty is estimated with adaptation and compensated. Meanwhile, the adaptive robust method is applied to deal with the non structured uncertainty by estimating unknown parameters in the upper bounding function. The modi®cation scheme [1] is used to cease parameter adaptation in accordance with the adaptive robust control law. The backstepping method [2] is also adopted in this chapter to deal with a system not in the parametric±pure feedback form, which is usually necessary for the application of backstepping control scheme. The new control scheme guarantees the uniform boundedness of the system and at the same time, the tracking error enters an arbitrarily designated zone in a ®nite time. The eectiveness of the proposed method is demonstrated by the application to PM synchronous motors.
309
12.1 Introduction Numerous adaptive robust control algorithms for systems containing uncertainties have been developed [1]±[11]. In [3] variable structure control with an adaptive law is developed for an uncertain input±output linearizable nonlinear system, where linearity-in-parameter condition for uncertainties is assumed. The unknown gain of the upper bounding function is estimated and updated by adaptation law so that the sliding condition can be met and the error state reaches the sliding surface and stays on it. To deal with a class of nonlinear systems with partially known uncertainties, in [4] an adaptive law using a dead zone and a hysteresis function is proposed to guarantee both the uniform boundedness of all the closed loop signals and uniform ultimate boundedness of the system states. In both control schemes, it is assumed that the system uncertainties are bounded by a bounding function which is a product of a set of known functions and unknown positive constants. The objective of adaptation is to estimate these unknown constants. In [1], a new adaptive robust control scheme is developed for a class of nonlinear uncertain systems with both parameter uncertainties and exogenous disturbances. Including the categories of system uncertainties in [3] and [4] as its subsets, it is assumed that the disturbances are bounded by a known upper bounding function. Furthermore, the input distribution matrix is assumed to be constant but unknown. In this chapter we proposed a continuous adaptive robust control scheme which is the extension of [1] in the sense that more general classes of nonlinear uncertain dynamical systems are under consideration. The unknown input distribution matrix of the system input can be state dependent here instead of being a constant matrix in [1]. To reduce the robust control gain and widen the application scope of adaptive techniques, the system uncertainties are supposed to be composed of two dierent categories: the ®rst can be separated and expressed as the product of known function of states and a set of unknown constants, and the other category is not separable but with partially known bounding functions. It is further assumed that the bounding function is convex to the set of unknown parameters, i.e. the bounding function is no longer linear in parameters. The ®rst category of uncertainties is dealt with by means of the well-used adaptive control method. Meanwhile an adaptive robust method is applied to deal with the second category of uncertainties, where the unknown parameters in the upper bounding function are estimated with adaptation. It should also be noted that the backstepping method [2] is adopted in this chapter to deal with a system not in the parametric±pure feedback form, which is usually necessary for the application of a backstepping control scheme. The proposed method is further applied to a permanent magnet synchronous (PMS) motor, which is a typical nonlinear control system. The dynamics of the PM synchronous motor can be presented by a dynamic electrical subsystem
310
and a mechanical subsystem, which are nonlinear dierential equations. Strictly speaking, most control methods for permanent magnet synchronous motors are only locally stable because the d-axis current is assumed to be zero and the design procedure is based on the reduced model. In this chapter, instead of only zeroing d-axis current, the extra d-axis control input voltage is used to deal with the nonlinear coupling part of the dynamics as well. This chapter is organized as follows. Section 12.2 describes the class of nonlinear uncertain systems to be controlled. Section 12.3 gives the design procedure of the adaptive robust control and the stability analysis. Section 12.4 describes the application of the proposed control method to the PM synchronous motors.
12.2
Problem formulation
Consider a class of uncertain dynamical system described by x_ 0 f 0 t B0 pg0 p t x1 g0 x p; !; tx2
12:1
x_ 1 f 1 x; t B1 pI E1 x; p; tu1 t g1 x; p; t g1 x; p; !; t 12:2 x_ 2 f 2 x; t B2 pI E2 x; p; tu2 t g2 x; p; t g2 x; p; !; t 12:3 where xi xi1 ; xi2 ; . . . ; xini ni ; i 0; 1; 2; are the measurable state vectors of the system, where n0 n1 and n0 n1 n2 n; x n is de®ned as
ni
x x
0 ; x1 ; x2 ; ui ui1 ; ui2 ; . . . ; uini , i 1; 2, are the control inputs of the system; p is an unknown system parameter vector and is the set of admissible system parameters; f i ni , i 0; 1; 2, are nonlinear function vectors; gi ni , i 0; 1; 2, and g0 n0 n2 , gi ni , i 1; 2, are nonlinear uncertain function vectors of the state x, unknown parameter p, time t as well as a set of random variables !. Here we make the following assumptions: (A1) f 0 t, f 1 x; t and f 2 x; t are known nonlinear function vectors. The matrices Bi p, i 0; 1; 2, are unknown but positive de®nite. (A2) For Ei ni ni , i 1; 2 t 0; x p
rmaxi 12 Ei 12 Ei rmini > 1
12:4
where indicates the eigenvalues of `'. (A3) The structured uncertainty gi ni , i 0; 1; 2, are nonlinear function vectors which can be expressed as
311
g0 p t 0 p0 t
g1 x; p; t 1 p1 x; t g2 x; p; t 2 p2 x; t
i diag
i1 ; . . . ; ini
i i1 ; i2 ; . . . ; ini ;
i 0; 1; 2
12:5
where 0 , 1 and 2 are unknown parameter matrices and 0 , 1 and 2 are known function vectors. The nonstructured uncertainty gi x; p; !; t, i 0; 1; 2, are bounded such that t x p gi x; p; !; t di x; qi ; t
12:6
where represents the Euclidean norm for vectors and the spectral norm for matrices; is a compact subset of n in which the solution of (12.1)±(12.3) uniquely exists with respect to the given desired state trajectory xd t. di x; qi ; t, i 0; 1; 2, are upper bounding functions with unknown parameter vectors qi . Here di x; qi ; t is dierentiable and convex to qi , that is
@d 12:7 di x; qi2 ; t d x; qi1 ; t qi2 qi1 @qi qi1 The control objective is to ®nd suitable control inputs u1 and u2 for the state x0 to track the desired trajectory xd t n0 , where xd is continuously dierentiable.
Remark 2.1 The sub system (12.1) has x1 as its input. However, it is not in the parametric±pure feedback form due to the existence of the nonlinear uncertain term g0 x; p; !; tx2 . Thus the well-used backstepping design needs to be revised to deal with the dynamical system (12.1)±(12.3). Remark 2.2 It should be noted that gi x; p; t can be absorbed into gi x; p; !; t. However, it is obviously more conservative. This can be clearly shown through the following example. Assume that the structured uncertainty is g 1 1 2 2 with actual values 1 a, 2 a and a is an unknown constant. Assume that the nonlinear function 2 1 , where 1 . Then g 1 1 2 2 1 , where
1 ; 2 a; a and 1 1 ; 2 . The upperbound parameter to be estimated is a; a . This implies that the actual uncertainty g a has been ampli®ed to its normed product a 1 a 1 , which is obviously much larger than a even if the estimates converge to the true values. On the contrary, if the uncertainty is expressed by (12.5), the unknown parameters to be
312
estimated is a a . This means that, when the estimated parameters are near the true values, the estimated uncertainty of g will be able to approach the actual uncertainty a.
12.3
Adaptive robust control with l-modi®cation
The adaptive robust technique is combined with backstepping method in this section to develop a controller which guarantees the global boundedness of the system. The design procedures are presented in detail as follows. De®ne the measured state tracking error vector as e0 x0 xd
12:8
and the parameter matrices as i XB1 i ; We further denote z0 e0 , z1 x1 auxiliary control xref 1 is de®ned as
i 0; 1; 2
xref 1 ,
z2 x2 , z
z
0 ; z1 ; z2 ,
^ ^ _ d0 xref 1 K0 z0 0 0 0 f 0 x
12:9 and the 12:10
^ 0 and ^ 0 are the estimates of 0 and 0 where K0 is a gain matrix. respectively. The ®rst order derivative of xref 1 is derived as follows: ^ 0 _0 ^_ 0 0 ^_ 0 f 0 x_ d ^ 0 f_0 x d0 _0 x_ ref 0 1 K0 z f 1 0 0
where
12:11
^ 0 _0 ^_ 0 0 ^_ 0 f 0 x_ d ^ 0 f_0 x d f 1 K0 f 0
f 0 f 0 x_ d
0 K0 B0 0 ; K0 B0
0 0 ; z1 xref 1
12:12
Then the plant (12.1)±(12.3) can be rearranged as follows: ref z_ 0 f 0 t B0 pg0 p; t g0 z; p; xref 1 ; !; tz2 z1 x1
12:13
ref ref z_ 1 f 1 z; xref 1 ; t B1 pI E1 z; p; x1 ; tu1 t g1 z; p; x1 ; t
g1 z; p; xref 1 ; !; t z_ 2
f 2 z; xref 1 ; t
B2 pI
g2 z; p; xref 1 ; !; t
12:14 E2 z; p; xref 1 ; tu2 t
g2 z; p; xref 1 ; t 12:15
where
313
f 1 f 1 f 1
g1 1 1
1 1 ; B1 1 0
1 1 ; 0
The adaptive robust control law.
12:16
De®ne the parameter error matrices as
~ qi qi ^qi ;
12:17
~ i i ^ i;
12:18
~ i i ^ i;
i 0; 1; 2
12:19
^ i; ^ i are the estimates of qi ; i ; i , i 0; 1; 2, respectively. where ^ qi ; The control law ui , i 1; 2, are chosen to be ui uc i uv i
12:20
^ i i ^ i f vd i 1vd uci Ki zi i 2z0 i 0 i uv i
r2maxi uci 2 zi 1 rmini rmaxi z
i uci "vi
vd i
^2di zi ^di zi "di
vd0
^2d0 z0 2 z2 ^d0 z0 z2 "d0
12:21
where Ki ni ni , i 1; 2, is a gain matrix; "vi , i 1; 2, and "di , i 0; 1; 2, are positive constants; f 2 Xf 2 ; ^di Xdi z; ^qi ; xref 1 ; t; The corresponding adaptive laws are de®ned as
where
ij ;
^_ i
i1 zi i
^_ i
i2 zi f i
^ i i1
^ i2 @ ^ q_ i i3 zi di i3 ^qi ; @qi ^qi
i 0; 1; 2
12:22
j 1; 2; 3 are positive de®nite matrices chosen to be ij
diag ij1 ; ij2 ; . . . ; ijni
12:23
314
ij ; j 1; 2; 3, which constitute the -modi®cation scheme, are de®ned as kij "0 z z E0 ij 12:24 0 elsewhere where kij , i 0; 1; 2, j 1; 2; 3, are positive constants. E0 Xe : e < "0
12:25
where "0 is a positive constant specifying the desired tracking error bound. Convergence analysis. following theorem.
For the above adaptive robust controller, we have the
Theorem 3.1 By properly choosing the control gain matrix, the proposed adaptive robust control law (12.20)±(12.24) ensures that the system trajectory enters the set E0 in a ®nite time. Moreover, the tracking errors as well as the parameter estimation errors are bounded by the set 2 X ~ i; ~ ~ i trace ~ ~ i; ~ ~ i ~q ~qi D z; qi ; i 0; 1; 2 : z z trace i i i i0
X 2
ki1 "0 trace
k " trace k " q q 2" < 12 k i i2 0 i i3 0 i i i i i0
12:26 where k is de®ned to be k max kij ;
i 0; 1; 2;
j 1; 2; 3
kij
1 k ij min min B1 i ; min
k ij
2 minmin K; kij 1 max max B1 i ; max ij
1 ij
max A and min A indicate the maximum and minimum eigenvalues of the matrix A respectively, and " and are positive values to be de®ned later. Proof
The following positive de®nite function V is selected
where
V V1 V2 V2
1 1 ~
V1 12 z
0 B0 z0 2 trace 0
1 ~ 01 0
~
12 trace 0
1 ~ 02 0
12 ~q
0
1 ~ 03 q0 ;
12:28
1 1 ~
V2 12 z
1 B1 z1 2 trace 1
1 ~ 11 1
~
12 trace 1
1 ~ 12 1
12 ~q
1
1 ~ 13 q1 ;
12:29
1 1 ~
V3 12 z
2 B2 z2 2 trace 2
1 ~ 21 2
~
12 trace 2
1 ~ 22 2
12 ~q
2
1 ~ 23 q2
12:30
12:27
315
Take the derivatives of V1 , V2 and V3 along the trajectory of the dynamic system (12.13)±(12.15), we have (See Appendices A±C)
1 ~ ~ ~ ~ ^d0 z0 z2 12 01 trace V_ 1 z
0 0 2 02 trace 0 0 0 K 0 z0 z0 z1 1
12 01 trace
0 0 2 02 trace 0 0
12:31
1 1 1 ~ ~ ~ ~ V_ 2 z
q
q1 1~ 1 K1 z1 z1 z0 2 11 trace 1 1 2 12 trace 1 1 2 13 ~
1
12 11 trace 1 1 12 12 trace
1 1 2 13 q1 q1 "v1 "d1 12:32
1 1 1 ~ ~ ~ ~ V_ 3 z
q
q2 2~ 2 K2 z2 z2 vd0 2 21 trace 2 2 2 22 trace 2 2 2 23 ~
1 1
12 21 trace
2 2 2 22 trace 2 2 2 23 q2 q2 "v2 "d2 12:33
Then we can easily get that V_ V_ 1 V_ 2 V_ 3 z Kz ^d0 z0 z2 z
2 vd 0
2 1X ~ ~ i i2 trace ~ ~ i i1 trace i i 2 i0
q
qi i1 trace
i3 ~ i ~ i i i2 trace i i i3 qi qi
z Kz ^d0 z0 z2
2 X i1
"vi "di
2 ^2d0 z0 2 z2 2 1X ~ ~ i i1 trace i ^d0 z0 z2 "d0 2 i0
~ ~ i i3 ~ q
qi i1 trace
i2 trace i i ~ i i i2 trace i i i3 qi qi
2 X i1
"vi "di
z Kz
2 1X ~ ~ i i2 trace ~ ~ i i3 ~q ~qi i1 trace i i i 2 i0
i1 trace
i i i2 trace i i i3 qi qi " P P where K diag K0 ; K1 ; K2 , and " 2i0 "di 2i1 "vi . By choosing K such that
min K
" c1 "20
12:34
12:35
where c1 is an arbitrary positive constant, then from (12.24) we have V_ z Kz " c1 ;
z n E0
12:36
Note that, in terms of the adaptive robust control law (12.20)±(12.24), V_ is a
316
continuous function. We can show that there exists a constant 0 < "0 < "0 such that (see Appendix E) V_ < 0
z n E0 ;
12:37
where E0 Xz : z < "0 is a subset of E0 . Noting the relation E0 E0 , (12.37) implies that the system will enter the set E0 in a ®nite time. When z E0 , it is obvious that kij "0 "0 ij kij "0 ;
i 0; 1; 2;
j 1; 2; 3
12:38
De®ne "0 "0 > 0, then from (12.24), (12.34) and (12.38) we obtain 1 V_ z Kz 2
2 X i0
~ ~ i ki2 trace ~ ~ i ki3 ~q ~qi ki1 trace i i i
ki1 "0 trace
i i ki2 "0 trace i i ki3 "0 qi qi "
k V
2 1X
ki1 "0 trace
i i ki2 "0 trace i i ki3 "0 qi qi " 2 i0
12:39
where k min k ij ; k ij
i 0; 1; 2;
j 1; 2; 3
2min min K; kij 1 max max B1 i ; max ij
By solving (12.39) we can establish that 2 1 X
k t ki1 "0 trace
Vt e Vt 0 i i ki2 "0 trace i i 2k i0 ki3 "0 q
i qi 2" ~ ~ and q converge exponentially to the residual set which implies that z, , 2 X ~ i; ~ ~ i trace ~ ~ i ~q ~qi ~ i; ~ D z; trace qi ; i 0; 1; 2 : z z i i i i0
k < 2
X 2 i0
ki1 "0 trace
i i
ki2 "0 trace
i i
ki3 "0 q
i qi
2"
317
where k max kij kij
i 0 1 2;
1 k ij min min B1 i ; min
j 1 2 3 1 ij
12.4 Application to PM synchronous motors Model of permanent magnet synchronous motor. A permanent magnet synchronous motor (PMSM) is described by the following subsystems: (1) a dynamic mechanical subsystem, which for the purposes of this discussion includes a single-link robot manipulator and the motor rotor; (2) a dynamic electrical subsystem which includes all of the motor's relevant electrical eects. d ! dt d! 1 Ld Lq Id f Iq Tsin dt J dId 1 ud !Lq Iq RId Ld dt
12:42
dIq 1 uq !Ld Id RIq !f Lq dt
12:43
12:40 12:41
where (12.40) and (12.41) present the dynamics of mechanical subsystem, and (12.42) and (12.43) are the dynamic electrical subsystem. In those equations, uq and ud are the input control voltages; Id and Iq are the motor armature current; R is the stator resistance; Ld and Lq are the self-inductances; J is the inertia angular momentum; and f is the ¯ux due to permanent magnet. For the above electromechanical model, we assume that the true states (i.e., , !, Id and Iq ) are all measurable. This model is obtained by using circuits theory principles and a particular dq reference frame. The control objective is to develop a link position tracking controller for the electromechanical dynamics of (12.40)± (12.43) despite parametric uncertainty. In this chapter we assume that all the motor parameters are unknown. Remark 4.1 In most existing control schemes for PM synchronous motors, the controllers are designed based on the following reduced model
318
d ! dt d! 1 f Iq T sin dt J dIq 1 uq RIq !f dt Lq
12:44 12:45 12:46
In this reduced model, component ud is adjusted to regulate current state Id and it is supposed that Id exactly equals to zero. Based on this reduced model, the position or velocity control is only directly related to the command voltage uq . Obviously, based on the reduced model, the control design will result in only a locally stable controller. Control Design. to be
For a given desired tracking state d t, de®ne a quantity z0 _ z0 ce e;
e d ;
c>0
12:47
where d t is at least twice continuously dierentiable. Dierentiating (12.47), multiplying by J and substituting the mechanical subsystem dynamics of (12.41), yields J z_0 Jce_ Ld Lq Id f Iq T sin J !_ d 12:48 Dividing (12.48) by f and rearranging terms yield
J z_0
1 ' 1 Iq L Id Iq
12:49
where 1 and '1 are de®ned as
1 J ; T ;
J , T and L are de®ned as J
J ; f
T
_ d ; sin '
1 ce_ ! T ; f
L
12:50
Ld Lq f
Denote z1 Iq Iqref , z2 Id , and the auxiliary reference current Iqref is de®ned as ^
Iqref 12:51 1 '1 k0 z0 where k0 is a positive constant gain. Then the PMSM model (12.40)±(12.43) can be transformed into ref z_0 J 1
1 '1 z1 Iq L Iq z2
12:52
z_1 L1 q uq 2 '2 3 '3
12:53
z_2 L1 d 4 '4 ud
12:54
where the unknown constant parameter vectors 2 , 3 and 4 and the known
319
regression vectors '2 , '3 and '4 are de®ned as (see Appendix E for the derivation of 2 , 3 , '2 and '3 ) f L T
2 Lq ; Lq ; Lq J J J ^11 Iq ; k1 c ^11 Id Iq ; k1 c ^11 sin '
2 k1 c
3 Ld ; R; f ; Lq _ d z0 '
'
3 !Id ; Id ; !; k1 ce_ ! 1
4
1 1 '1
A!
R; Lq
'
4 Id ; !Iq ^11 c!_ d ! d ^12 ! cos . Comparing (12.52)±(12.54) with where A! (12.13)±(12.15), it is obvious that PMSM dynamics is a particular case of the general nonlinear uncertain dynamic system (12.13)±(12.15). Therefore all the previous design and analysis can be applied directly. Note that in practical applications, the parameters J and T are constants but may vary in a wide range due to the variation of payload. On the other hand, the unknown motor parameters have fewer deviations from its rated values (nominal values) in comparison with that of load. Therefore, it would be more appropriate for us to deal with the unknown parameters J and T by using adaptive techniques and treat the bounded motor parameters by using robust methods. In this way, referring to (12.20)±(12.24) the control inputs with the corresponding adaptive laws are given as ^
uq k1 z1 z0 2 ' 2 vd 1 ud k2 z2 vd0 vd2 ^1 ^_ 1 '1 z0 1 ^_ 2 '2 z1 2 ^2 vd 0
L2max Iq 2 z0 2 z2 Lmax Iq z0 z2 "d0
2max3 '3 2 vd 1 z1 max3 '3 z1 "d1 vd 2
2max4 '4 2 z2 max4 '4 z2 "d2
12:55
where L Lmax , 3 max3 and 4 max4 . i , i 1; 2, are chosen to
320
be i
"0 z 0
z "0
elsewhere
12:56
where z z0 ; z1 ; z2 . Simulation Study. The permanent magnet synchronous motor with the following parameters is used to demonstrate the control performance Ld 25:0 103 H; f 0:90 N m=A;
Lq 30:0 103 H; R 5:0 ;
J 1:625 103 Kg m2 ;
T 2:2816 Kg A m=N s2
1 0:0018; 2:535 The following control parameters c 5:0;
"0 0:1;
"d0 "d1 "d2 0:5;
k1 k2 k3 8:0
are used for the proposed control scheme. The desired trajectory is chosen as 3 1 e0:1t sin t d 2 5 The initial states are 0 0:1 rad, !0 0:1 rad/sec, Id 0 0:05 A, Iq 0 0:01 A. The initial values of all the parameter estimates are set zero. The tracking error is shown in Figure 12.1. We can see that the proposed control method yields very good tracking performance. Meanwhile, the control inputs ud and uq are very smooth as shown in Figure 12.2. Figure 12.3 (a) shows the corresponding armature current Iq compared with auxiliary reference current Iqref while Id is shown in Figure 12.3 (b). It is easy to see that the current Iq asymptotically converges to Iqref .
12.5
Conclusion
In this chapter, an adaptive robust control scheme is developed for a class of uncertain systems with both unknown parameters and system disturbances. The uncertainties are assumed to be composed of two categories: the structured category and the nonstructured category with partially known bounding functions. The structured uncertainty is estimated with the adaptive method. Meanwhile, the adaptive robust method is applied to deal with the nonstructured uncertainty, where the unknown parameters in the upper bounding function are estimated with adaptation. It is shown that the control scheme developed here can guarantee the uniform boundedness of the system and assure that the tracking error enters the arbitrarily designated zone in a ®nite time. The eectiveness of the control scheme is veri®ed by theoretical analysis, as well as applications to a permanent magnet synchronous motor.
321
a e
time (sec) b !
time (sec)
Figure 12.1
e !
Appendix A
The derivative of V1
1 ~
_ 0 trace V_ 1 z
0 B0 z
_ 1 ^ 01
~
trace 0
1 ^_ 02 0
~q
0
1 _ q0 03 ^
1 ref ~
z
0 B0 f 0 B0 0 0 g0 z2 z1 x1 trace 0
~
trace 0
_ 1 ^ 02 0
~ q
0
_ 1 ^ 01 0
1 _ q0 03 ^
ref
~
^ z
0 0 0 z1 x1 0 f 0 g0 z2 trace 0 z0 0 01 0
~ z0 f 02 ^ 0 ~ trace q
0 0 0
1 _ q0 03 ^
~
~
~ ~ z
0 K0 z0 z0 0 0 z1 0 f 0 g0 z2 z0 0 0 z0 0 f 0 @d ~ ^ ~ ^ q
q0 z0 q0 ^ q0 0 01 trace 0 0 02 trace 0 0 03 ~ 0^ @q 0 ^q0
z
0 K0 z0
z
0 z1
d0 z0 z2 z0 d0 ^d0
~ ^ ~ ^ q
q0 01 trace 0 0 02 trace 0 0 03 ~ 0^
~ 0 ~ 0 ^d0 z0 z2 01 trace z
0 K0 z0 z0 z1 0
~ 0 ~ 0 03 ~q q0 ~q0 02 trace 0 0
322
a uq
time (sec) b ud
time (sec)
Figure 12.2
uq
ud
~ ~
~ ^d0 z0 z2 01 trace z
0 K 0 z0 z0 z1 0 0 01 trace 0 0
~ ~ ~
02 trace q
q0 03 ~q
0 0 02 trace 0 0 03 ~ 0~ 0 q0
1 ~ ~ ~ ~ ^d0 z0 z2 12 01 trace z
0 K 0 z0 z0 z1 0 0 2 02 trace 0 0 1 1
q
q0 12 01 trace
12 03 ~ 0 0 2 02 trace 0 0 2 03 q0 q0 0~
A:1
In above derivation the following property of trace is used: trace Q vw v Qw
A:2
where v n1 , w l1 , and Q nl .
Appendix B
The derivative of V2
1 ~
_ 1 trace V_ 2 z
1 B1 z 1
_ 1 ^ 11 1
~
trace 1
_ 1 ^ 12 1
~q
1
1 ~
z
1 B1 f 1 B1 I E1 u1 1 1 g1 trace 1
~
trace 1
_ 1 ^ 12 1
~ q
1
1 _ q1 13 ^
1 _ q1 13 ^
_ 1 ^ 11 1
323
a Iq and I ref q
time (sec) b
Id
time (sec)
Figure 12.3
Iq
I ref q
Id z
1 1 1 I E1 uc1 uv1 g1 1 f 1
~ z1 ^ ~
^ trace q
11 1 trace 1 z1 f 1 12 1 ~ 1 1 1
1 _ q1 13 ^
~ ~ z
1 K1 z1 z0 1 1 1 f 1 I E1 uv1 E1 uc1 vd1 g1
~ ~ ^ ~ ~ ^ q
z
1 1 1 z1 1 f 1 11 trace 1 1 12 trace 1 1 ~ 1
1 _ q1 13 ^
z
1 K1 z1 z1 z0 d1 z1 z1 vd1 rmax1 z1 uc1
z
1 I E1
r2max1 uc1 2 z1 ~ ^ 11 trace 1 1 1 rmin1 rmax1 z
u " c v 1 1 1
~ ^ 12 trace q
1 1 ~ 1
1 _ q1 13 ^
A:3
Using the fact that
z
i I Ei zi zi I
I 12 Ei 12 Ei
1; 1 rmini
Ei Ei
zi 2
i 1; 2
A:4
324
it follows that
V_ 2 z
1 K 1 z1 z1 z0
^2d1 z1 2
@d1 ^ d1 z1 z1 q1 q1 @qi ^q1 ^d1 z1 "d1
~ ^ ~ ^ q
q1 "v1 11 trace 1 1 12 trace 1 1 13 ~ 1^
z
1 K 1 z1 z1 z0
^2d1 z1 2 d1 z1 z1 d1 ^d1 ^d1 z1 "d1
~ ^ ~ ^ 11 trace q
q1 "v1 1 1 12 trace 1 1 13 ~ 1^
~ ^ ~ ^ z
1 K1 z1 z1 z0 11 trace 1 1 12 trace 1 1
q
q1 "v1 "d1 13 ~ 1^
1 1 1 ~ ~ ~ ~ q
q1 z
1~ 1 K1 z1 z1 z0 2 11 trace 1 1 2 12 trace 1 1 2 13 ~
1
12 11 trace 1 1 12 12 trace
1 1 2 13 q1 q1 "v1 "d1
Appendix C
A:5
The derivative of V3
1 ~
_ 2 trace V_ 3 z
2 B2 z 2
_ 1 ^ 21 2
~
trace 2
_ 1 ^ 22 2
~q
2
1 ~
z
2 B2 f 2 B2 I E2 u2 2 2 g2 trace 2
~
trace 2
_ 1 ^ 22 2
~ q
2
1 _ q2 23 ^
_ 1 ^ 21 2
1 _ q2 23 ^
z
2 2 2 I E2 uc2 uv2 g2 2 f 2
~ z2 ^ ~
^ q
trace 21 2 trace 2 z2 f 2 22 2 ~ 2 2
1 _ q2 g 23 ^
~ ~ z
2 K2 z2 2 2 2 f 2 I E2 uv2 E2 uc2 vd0 vd2 g2
~ ~ ^ ~ ~ ^ q
z
2 2 2 z2 2 f 2 21 trace 2 2 22 trace 2 2 ~ 2
1 _ q2 23 ^
z
2 K2 z2 z2 vd0 z2 vd2 d2 z2 rmax2 z2 uc2
z
2 I E2
r2max2 uc2 2 z2 ~ ^ 21 trace 2 2 1 rmin2 rmax2 z
2 uc2 "v2
^ ~ 22 trace q
2 2 ~ 2
1 _ q2 23 ^
A:6
325
Using the fact of (A.6) it follows that
V_ 3 z
2 K2 z2 z2 vd0
^2d2 z2 2 @d d2 z2 z2 q2 ^q2 2 @qi ^q2 ^d2 z2 "d2
~ ^ ~ ^ q
q2 "v2 21 trace 2 2 22 trace 2 2 23 ~ 2^
z
2 K2 z2 z2 vd0
^2d2 z2 2 d2 z2 z2 d2 ^d2 ^d2 z2 "d2
~ ^ ~ ^ q
q2 "v2 21 trace 2 2 22 trace 2 2 23 ~ 2^
~ ^ ~ ^ q
q2 "v2 "d2 z
2 K2 z2 z2 vd0 21 trace 2 2 22 trace 2 2 23 ~ 2^
1 1 1 ~ ~ ~ ~ q
q2 z
2~ 2 K2 z2 z2 vd0 2 21 trace 2 2 2 22 trace 2 2 2 23 ~ 1 1
12 21 trace
2 2 2 22 trace 2 2 2 23 q2 q2 "v2 "d2
A:7
Appendix D From (12.34), it can be seen that V_ < 0 if z Kz >
2 1X
i1 trace
i i i2 trace i i i2 qi qi 2" 2 i0
A:8
then "0 can be easily determined by solving the following equation 2 " c1 1X 2
i1 trace
z i i i2 trace i i i2 qi qi 2" 2 i0 "20
Substituting ij in terms of (12.24) and letting z "0 yields 2 " c1 "0 2 1 X 1
ki1 "0 "0 trace
i i 2 ki2 "0 "0 trace i i 2 2 i0 "0
Denote
12 ki3 "0 "0 q
i qi 2"
a
" c1 "20
b
2 1X
ki1 trace
i i ki2 trace i i ki3 qi qi 2 i0
c
2 1X
ki1 "0 trace
i i ki2 "0 trace i i ki3 "0 qi qi 2" 2 i0
A:9
326
then equation (A.9) can be transformed to the following 2
a"0 b"0 c 0 The solutions of above equation are "0
b
b2 4ac 2a
A:10
It is obvious that the solutions . Notice that "0 is positive, hence the desired solution is b b2 4ac "0 >0 A:11 2a
Appendix E
De®nitions of 2 , 3 , '2 and '3
Dierentiating z1 , yields z_ 1 I_q I_qref
A:12
From (12.51), we have I_qref z0
1 '1
1 '
1
1 '1
^
_ 1 k1 z_ 0 '1 1' ^
_1 1'
k1
'1 Iq L Id Iq J 1
A:13
^
_ 1 , it follows For 1' _1 ^11 c!_ !_ d ! d ^12 ! cos ^
1'
^11 c ^11 c!_ d !d ^12 ! cos Ld Lq Id Iq f Iq T sin J
0 '0 A!
A:14
where 0 , '0 and A! are de®ned by Ld Lq f T
; 0 ; J J J
A:15
^11 Id Iq ; Iq ; sin '
0 c
A:16
^11 c!_ d ! d ^12 ! cos A!
A:17
327
Multiplying (A.12) by Lq and substituting (12.43), we obtain Lq z_1 uq !Ld Id RId !f k1
1
Lq 1 '
' ' A ' I L I I 1 ! 1 q d q 1 1 0 0 J 1
uq
2 ' 2 3 ' 3
A:18
where
2
f L T Lq ; Lq ; Lq J J J
^11 Iq ; k1 c ^11 Id Iq ; k1 c ^11 sin '
2 k1 c
3 Ld ; R; f ; Lq _ d 1 '
'
3 !Id ; Id ; !; k1 ce_ ! 1
1 1 '1
A!
References [1] Jian-Xin Xu, Tong-Heng Lee and Qing-Wei Jia, (1997) `An adaptive Robust Control Scheme for a Class of Nonlinear Uncertain Systems', International Journal of Systems Science, Vol. 28, No. 4, 429±434. [2] Kokotovic, P. V., (1992) `The Joy of Feedback: Nonlinear and Adaptive', IEEE Control systems, Vol. 12, No.6, 7±17. [3] Teh-Lu Liao, Li-Chen Fu and Chen-Fa Hsu, (1990) `Adaptive Robust Tracking of Nonlinear Systems and With an Application to a Robotic Manipulator', Systems & Control Letters, Vol. 15, 339±348. [4] Brogliato, B. and Tro®no Neto, A. (1995) `Practical Stabilization of a Class of Nonlinear Systems with Partially Known Uncertainties', Automatica, Vol. 31, No. 1, 145±150. [5] Ioannou, P. A. and Jing Sun, (1996) Robust Adaptive Control. Prentice-Hall, Inc. [6] Ioannou, P. A. and KokotovicÂ, P. V. (1983) Adaptive Systems with Reduced Models. Springer-Verlag, Berlin Heidelberg. [7] Peterson, B. B. and Narendra, K. S. (1982) `Bounded Error Adaptive Control', IEEE Transac. Autom. Contr., Vol. 27, No. 6, 1161±1168. [8] Taylor, D. G. KokotovicÂ, P. V. Marino, R. and Kanellakopoulos, I. (1989) `Adaptive Regulation of Nonlinear Systems with Unmodeled Dynamics', IEEE Trans Autom. Control, Vol. 34, No. 1, 405±412. [9] Sastry, S. S. and Isidori, A. (1989) `Adaptive Control of Linearizable Systems', IEEE Transac. Autom. Contr., Vol. 34, No. 11, 1123±1131. [10] Narendra, K. S. and Annaswamy, A. M. (1989) Stable Adaptive Systems. PrenticeHall, Englewood Clis, NJ. [11] KokotovicÂ, P. V. (1991) Foundations of Adaptive Control. Edited by Thomam, M. and Wyner, A. Springer-Verlag, CCES.
Index
Active identi®cation scheme, 159, 161, 165, 166±78 ®nite duration, 178 input selection for identi®cation, 173±8, 179, 180 pre-computation of projections, 168±73, 180 orthogonalized projection algorithm, 169±70 orthogonalized projection for output-feedback systems, 170±1 pre-computation procedure, 1713 See also Discrete-time nonlinear systems Actuator compensation, adaptive inverse approach, 260±85 designs for multivariable systems, 275±80 nonlinearity model and its inverse, 277 output feedback designs, 280 state feedback designs, 278±80 designs for nonlinear dynamics, 281±4 output feedback designs, 271±5, 280 model reference design, 272 PID design, 273±5 pole placement design, 272±3
output feedback inverse control, 269±71 adaptive law, 271 error model, 270±1 reference system, 269±70 parametrized inverses, 263±5 control error, 264±5 inverse examples, 264 inverse model, 263 plants with actuator nonlinearities, 262±3 dead-zone example, 262±3 nonlinearity model, 262 state feedback designs, 265±9, 278±80 continuous-time Lyapunov design, 268±9 error model, 265±6 gradient projection adaptive law, 266±8 Adaptive backstepping, 134±5, 281 with tuning functions, 120, 127, 193±7 See also Backstepping methodology Adaptive control algorithms, 23±5 robustness problems, 81 Adaptive control Lyapunov functions (aclf), 185, 188 Adaptive internal model control schemes, 7±21
330
Index
certainty equivalence control laws, 9± 12, 15±16 adaptive H2 optimal control, 11, 13± 14, 18±19 adaptive H optimal control, 11±12, 13, 14, 18±19 model reference adaptive control, 10±11, 13, 14, 18 partial adaptive pole placement, 10, 13, 14, 18 design of robust adaptive law, 7±9 direct adaptive control with dead zone, 27±8 robustness analysis, 12±18 simulation examples, 18±19 stability analysis, 12±18 See also Individual model control schemes Adaptive nonlinear control, see Nonlinear systems Adaptive passivation, 119±21, 123±4, 127±38 adaptively feedback passive (AFP) systems, 128±30, 132 examples and extensions, 135±8 adaptive output feedback passivation, 136±8 controlled Dung equations, 135±6 recursive adaptive passivation, 131±5 Adaptive robust control, see Robust adaptive control Adaptive stabilizability, 127±8 Adaptive stabilization, 81±2 direct localization principle, 89±101 localization in the presence of unknown disturbance, 99±101 optimal localization, 94±9 indirect localization principle, 101±8 localization in the presence of unknown disturbance, 108 See also Switching control Adaptive tracking control Lyapunov function (atclf), 184, 186, 188±93, 201
Adaptive variable structure control, 41±61 case of arbitrary relative degree, 49±56 robustness, 52±5 stability, 52±5 tracking performance, 52±6 case of relative degree one, 46±9 robustness, 47±8 stability, 47±8 tracking performance, 47±9 computer simulations, 56±9 problem formulation, 43±6 MRAC based error model, 44±6 plant description and control objectives, 43±4 Adaptively feedback passive (AFP) systems, 128±30, 132 Artstein, Z., 188 Averaging ®lter, 50 Backlash, 260±2, 274±5, 277 inverse, 264, 274±5 See also Actuator compensation Backstepping methodology, 160 adaptive backstepping, 134±5, 281 with tuning functions, 120, 127, 193±7 for system not in the parametricpure feedback form, 308, 309 inverse optimality via backstepping, 202±5 Basar, T., 186 Blaschke product, 4±5, 11 Bounded-input bounded-output (BIBO) stable systems, 125 Bounded-input bounded-state (BIBS) stable systems, 132 Byrnes, C.I., 123, 136 Certainty equivalence control laws, 9±12, 15±16, 72, 185, 216 adaptive H2 optimal control, 11, 13±14, 18±19 adaptive H1 optimal control, 11±12, 13, 14, 18±19
Index partial adaptive pole placement, 10, 13, 14, 18 See also Model reference adaptive control (MRAC) Chattering behaviour, 56 Cholesky decomposition, 69, 71 Concave covers, 223±8 de®nitions, 223 examples of, 224±8 properties of, 223±4 Concave functions, 217, 221±3 de®nitions, 221 properties of, 221±3 Continuous-time Lyapunov design, 268±9 Control Lyapunov function (clf), 188 Controlled Dung equations, 135±6 Converging-output converging-state (COCS) function, 128, 132 Converse Lyapunov Theorem, 130 Convex functions, 217, 221±3 de®nition, 221 properties of, 221±3 Dead zone based methods, 23±5, 260±2, 267, 277 inverse of dead zone characteristic, 264 ordinary direct adaptive control with dead zone, 27±8 robust direct adaptive control, 28±31 See also Actuator compensation Deadbeat control strategy, 67, 160, 163, 166 Dense search method, 83±4 Didinsky, G., 186 Direct adaptive control, 28±31 multi-input multi-output systems, 289±96 with dead zone, 27±8 Direct localization principle, 89±101 localization in the presence of unknown disturbance, 99±101 optimal localization, 94±9 Discrete-time nonlinear systems: active identi®cation, 159±82
331
avoidance of nonlinear parametrization, 164±6 ®nite duration, 178 input selection for identi®cation, 173±8, 179, 180 pre-computation of projections, 168±73, 180 problem formulation, 161±6 second-order example, 162±4 adaptive stabilization, 80±115, 245 direct localization principle, 89±101 indirect localization principle, 101±8 problem statement, 86±9 simulation examples, 108±12 See also Nonlinear systems Dung equations, 135±6 Dynamic feedback, 187, 204, 209 Dynamic normalization technique, 151±5 Feedback passivation, see Adaptive passivation Finite escape time, 162, 166 Freeman, R.A., 185 Friction models, 236±40 Fuzzy systems, 287 applications, 299±305 two-link robot arm, 301±5 direct adaptive control, 290±2 indirect adaptive control, 297±9 Globally adaptively stabilizable systems, 128, 129 Globally uniformly asymptomatically stable (GUAS) systems, 124 Gradient projection adaptive law, 266±8 H2 optimal control: adaptive internal model control schemes, 11, 13±14 simulation example, 18±19 nonadaptive control schemes, 4±5 H1 optimal control: adaptive internal model control schemes, 11±12, 13, 14 simulation example, 18±19 nonadaptive control schemes, 5±6
332
Index
Hamilton±Jacobi±Bellman (HJB) pde, 184, 185 Hurwitz polynomial of degree, 7±8, 9 Hysteresis, 260±2, 277 inverse, 264 See also Actuator compensation Hysteresis switching, 11, 84 Identi®cation scheme, see Active identi®cation scheme Indirect adaptive control, 63±79 adaptive control law, 69±74 properties of identi®cation scheme, 69±70 strategy, 70±4 adaptive control scheme, 68±9 multi-input multi-output systems, 296±9 problem formulation, 65±7 simulation examples, 74±8 system with noise, 75±8 system without noise, 74±5 Indirect localization principle, 101±8 localization in the presence of unknown disturbance, 108 Input-to-output practically stable (IOpS) systems, 125, 126 Input-to-output stable (IOS) systems, 125 Input-to-state practically stable (ISpS) systems, 125, 126 Input-to-state stable (ISS) systems, 120, 124±5 Internal model control (IMC) schemes, 1±2 IMC ®lter, 5, 11 known parameters, 2±7 H2 optimal control, 4±5 H1 optimal control, 5±6 model reference control, 3±4 partial pole placement control, 3 robustness to uncertainties, 6±7 See also Adaptive internal model control schemes Inverse optimal control, 185±6 inverse optimal adaptive tracking, 197± 202
inverse optimality via backstepping, 202±5 See also Actuator compensation, adaptive inverse approach Ioannou, P.A., 201 Isidori, A., 123, 136 Kharitonov Theory, 10 Kokotovic, P.V., 185, 186 Krstic, M., 185, 186, 196±7, 205±8, 210 Kuhn Tucker theorem, 249 Lagrangian function, 249 LaSalle's theorem, 191 Levy±Desplanques theorem, 298 Linear-quadratic-regulator (LQR), 201 Linearly parametrized systems, 215±16 Localization method, 80, 85 direct localization principle, 89±101 localization in the presence of unknown disturbance, 99±101 optimal localization, 94±9 indirect localization principle, 101±8, 111±12 localization in the presence of unknown disturbance, 108 simulation examples, 108±12 constant parameters, 109±10 indirect localization, 111±12 parameter jumps, 110±11 LU decomposition, 69 Magnetic bearing system, 242±5 adaptive control based on nonlinear parametrization, 243±5 Martensson, B., 81±3 Min-max optimization, 215, 217, 245 Model reference adaptive control (MRAC), 10±11, 41 output feedback design, 272 simulation example, 18 stability and robustness analysis, 13, 14 variable structure MRAC system, 41±3 MRAC based error model, 44±6 See also Adaptive variable structure control
Index Model reference control, nonadaptive control schemes, 3±4 Morse, A.S., 84 Multi-input multi-output (MIMO) systems, 287±306 direct adaptive control, 289±96 indirect adaptive control, 296±9 See also Fuzzy systems Multivariable systems, 275±80 nonlinearity model and its inverse, 277 output feedback designs, 280 state feedback designs, 278±80 Neural networks, 216, 287 See also Fuzzy systems Newton±Raphson method, 227 Nonadaptive control schemes, 2±7 H2 optimal control, 4±5 H1 optimal control, 5±6 model reference control, 3±4 partial pole placement control, 3 robustness to uncertainties, 6±7 See also Internal model control (IMC) schemes Nonlinear gain, 124±7 input-to-state stable (ISS) systems, 124±5 nonlinear small gain theorems, 125±7 Nonlinear parametrization (NP), 216±18, 258, 277 avoidance of, 164±6 active identi®cation, 165 orthogonalized projection estimation, 166 projection measurements, 165 regressor subspace, 164±5 concave covers, 223±8 de®nitions, 223 examples of, 224±8 properties of, 223±4 concave/convex functions, 220±3 de®nitions, 221 properties of, 221±3 problem statement, 218±20 stable adaptive NP systems, 228±35
333
adaptive controller, 230±4 adaptive observer, 234±5 extensions to the vector case, 229 low-velocity friction model, 236±40 magnetic bearing system, 242±5 new error model, 228±9 scalar error model with linear parameters, 229±30 stirred tank reactors (STRs), 240±2 structure of adaptive controller, 218±20 See also Nonlinear systems Nonlinear systems: adaptive robust control for uncertain dynamical systems, 308±27 adaptive robust control with modi®cation, 312±17 application to PM synchronous motors, 317±20 problem formulation, 310±12 nonsmooth nonlinearities, 260±1 designs for, 281±4 See also Actuator compensation, adaptive inverse approach optimal adaptive tracking, 184±213 adaptive backstepping, 193±7 adaptive tracking control Lyapunov function (atclf), 188±93 design for strict-feedback systems, 205±9 inverse optimal adaptive tracking, 197±202 inverse optimality via backstepping, 202±5 problem statement, 186±8 transient performance, 209±10 See also Adaptive passivation; Discrete-time nonlinear systems; Nonlinear parametrization; Small gain techniques Nonsmooth nonlinearities, 260±1 designs for, 281±4 See also Actuator compensation, adaptive inverse approach; Nonlinear systems Nussbaum, R.D., 81±2
334
Index
Orthogonalized projection scheme, 160, 166 algorithm, 169±70 for output-feedback systems, 170±1 Output feedback systems, 271±5 adaptive output feedback passivation, 136±8 for multivariable systems, 280 inverse control, 269±71 adaptive law, 271 error model, 270±1 reference system, 269±70 model reference design, 272 orthogonalized projection for, 170±1 parametric-output-feedback form, 161±2 PID design, 273±5 pole placement design, 272±3 Overparametrization, 216 Pan, Z., 186 Parametric strict-feedback system, 139, 185±6, 205±9, 281 Parametric-output-feedback form, 161±2 Parseval's Theorem, 4 Partial pole placement control: adaptive internal model control schemes, 10, 13, 14 simulation example, 18 nonadaptive control schemes, 3 Passivation, 120, 123 See also Adaptive passivation Passivity, 121±3 strict passivity, 122 Pendulum example, 147±51 Permanent magnet synchronous (PMS) motor example, 309±10, 317±20 control design, 318±20 model of PMS motor, 317±18 simulation study, 320 PID design, 273±5 Piecewise-linear characteristic, 260±2, 277 inverse, 264 See also Actuator compensation Polar decomposition, 65, 69
Pole placement design, 272±3 Polynomial dierential operators, 81 Recursive adaptive passivation, 1315 Riccati equations, 184, 185, 201 Robots, 289 two-link robot arm, 301±5 Robust adaptive control: for nonlinear uncertain dynamical systems, 308±27 adaptive robust control with modi®cation, 312±17 application to PM synchronous motors, 317±20 problem formulation, 310±12 with less prior knowledge, 23±38 direct adaptive control, 28±31 problem formulation, 25±7 simulation example, 36±8 with least prior knowledge, 32±6 Robustness, 261 adaptive internal control schemes, 9± 10 adaptive algorithms, 81 robust adaptive law design, 7±9, 313±14 robustness analysis, 12±18 adaptive variable structure control, 47± 8, 52±5 nonadaptive control schemes, 6±7 Saturation function, 221 Self-excitation technique, 64 Separable sets, 95±6 Separation Principle, 148 Sepulchre, R., 185 Singular value decomposition (SVD), 65, 69 Small control property, 190 Small gain techniques, 6±7, 119±20, 138± 55 adaptive controller design, 139±45 initialization, 139±40 recursive steps, 140±4 small gain design step, 144±5 class of uncertain systems, 138±9
Index examples and discussions, 14755 pendulum example, 14751 robusti®cation via dynamic normalization, 151±5 nonlinear small gain theorems, 125±7 stability analysis, 146±7 Smooth control law, 203 Sontag, E.D., 188, 189±90, 196, 200, 202 Stability analysis, 12±18 adaptive variable structure control, 47±8, 52±5 small gain approach, 146±7 Stabilization, 185 See also Adaptive stabilization Stabilizing sets, 101±2 State feedback designs, 265±9, 278±80 continuous-time Lyapunov design, 268±9 error model, 265±6 gradient projection adaptive law, 266±8 multivariable systems, 278±80 Stirred tank reactors (STRs), 240±2 adaptive control based on nonlinear parametrization, 241±2 Strict passivity, 122 Strict-feedback systems, 139, 185±6, 205± 9 Sun, J., 201 Supervisory control, 84, 89 Switching control, 80±6 algorithms, 81 conventional switching control, 88±9 direct localization principle, 89±101 localization in the presence of unknown disturbance, 99±101 optimal localization, 94±9
335
indirect localization principle, 101±8 localization in the presence of unknown disturbance, 108 problem statement, 86±9 simulation examples, 108±12 constant parameters, 109±10 indirect localization, 111±12 parameter jumps, 110±11 supervisory switching control, 84, 89 Sylvester resultant matrix, 64±5 Takagi±Sugeno fuzzy systems, see Fuzzy systems Tracking performance, adaptive variable structure control, 47±9, 52±6 See also Nonlinear systems Trajectory initialization, 210 Tuning functions, 120, 186, 191, 233 adaptive backstepping with, 120, 127, 193±7 Unboundedness observable (UO) function, 128 Uncertain discrete-time systems, see Discrete-time nonlinear systems Variable structure design (VSD), 41±2 See also Adaptive variable structure control Willems, J.C., 123 Youla parameter, 3 Young's inequality, 137 Zero-state detectability, 122